Engineering 27 (2023) 23–30

Contents lists available at ScienceDirect

Engineering
journal homepage: [Link]/locate/eng

Research
Smart Process Manufacturing toward Carbon Neutrality—Perspective

Active Machine Learning for Chemical Engineers:

A Bright Future Lies Ahead!
Yannick Ureel, Maarten R. Dobbelaere, Yi Ouyang, Kevin De Ras, Maarten K. Sabbe, Guy B. Marin,
Kevin M. Van Geem ⇑
Laboratory for Chemical Technology, Department of Materials, Textiles and Chemical Engineering, Ghent University, Ghent 9052, Belgium

a r t i c l e i n f o a b s t r a c t

Article history: By combining machine learning with the design of experiments, thereby achieving so-called active
Received 9 September 2022 machine learning, more efficient and cheaper research can be conducted. Machine learning algorithms
Revised 7 December 2022 are more flexible and are better than traditional design of experiment algorithms at investigating pro-
Accepted 28 February 2023
cesses spanning all length scales of chemical engineering. While active machine learning algorithms
Available online 1 August 2023
are maturing, their applications are falling behind. In this article, three types of challenges presented
by active machine learning—namely, convincing the experimental researcher, the flexibility of data cre-
Keywords:
ation, and the robustness of active machine learning algorithms—are identified, and ways to overcome
Active machine learning
Active learning
them are discussed. A bright future lies ahead for active machine learning in chemical engineering,
Bayesian optimization thanks to increasing automation and more efficient algorithms that can drive novel discoveries.
Chemical engineering Ó 2023 THE AUTHORS. Published by Elsevier LTD on behalf of Chinese Academy of Engineering and
Design of experiments Higher Education Press Limited Company. This is an open access article under the CC BY-NC-ND license
([Link]

1. Introduction machine learning as a supervised machine learning technique in

which the learner—that is, the machine learning model—is in con-
Experiments performed under well-defined conditions and cal- trol of the data from which it learns. In active machine learning,
culations based on first principles constitute the basis of engineer- machine learning algorithms are used to iteratively determine
ing research. In chemical engineering, these activities are aimed at, new experimental data, the so-called training data, based on
for example, the development and optimization of catalysts, reac- uncertainty criteria. It should be noted that ‘‘experimental” can
tion conditions, and reactor configurations. In the chemical indus- also refer to computationally expensive high-level simulations,
try, 51 billion USD was spent in 2017 on research and development such as high-level ab initio calculations of molecular properties
[1]. This illustrates the importance of high-quality data; however, or large eddy simulations of reactive flow with computational fluid
obtaining accurate data is tedious and error prone. The design of dynamics (CFD) codes [7]. Active machine learning consists of two
experiments (DoE) can help in extracting the maximal information branches with two different purposes: active learning and
with a minimum of effort [2,3], making sure that time and Bayesian optimization. Active learning aims to explore and model
resources are spent efficiently. By integrating machine learning a process with a minimum number of ‘‘experiments” to ensure
with DoE, a more flexible and efficient DoE is achieved. This so- accurate predictions over the entire design space [8]. Bayesian
called ‘‘active machine learning” allows a more effective selection optimization is essentially a machine learning-based optimization
of experimental conditions, particularly for high-dimensional and strategy, where iteratively new experimental data is selected to
highly nonlinear phenomena [4]. find an experiment that optimizes the objective [9]. Either active
Machine learning can facilitate the automation of the whole learning or Bayesian optimization can be employed for experimen-
experimental cycle, from experimental selection to model building tal selection, depending on whether the goal is to model a process
and data analysis [5]. While the most common field of application and acquire process knowledge or to optimize an objective.
in machine learning is model building and data analysis, the focus
of this article is on the potential of combining DoE with machine 1.1. Basic principles of active machine learning
learning for active machine learning. Olsson [6] defined active
Fig. 1 [10] illustrates the general workflow of active machine
⇑ Corresponding author. learning algorithms, starting with the initialization followed by
E-mail address: [Link]@[Link] (K.M. Van Geem). an iterative loop consisting of three phases. The critical first step

[Link]
2095-8099/Ó 2023 THE AUTHORS. Published by Elsevier LTD on behalf of Chinese Academy of Engineering and Higher Education Press Limited Company.
This is an open access article under the CC BY-NC-ND license ([Link]
Y. Ureel, M.R. Dobbelaere, Y. Ouyang et al. Engineering 27 (2023) 23–30

Fig. 1. Overview of the general active machine learning workflow, depicting initialization and iterative query selection. Reproduced from Ref. [10] with permission.

of initialization consists of clearly defining the research problem as After initialization, the active machine learning procedure con-
either the modeling of an output (active learning) or the optimiza- sists of three phases: the training of the machine learning model,
tion of an objective (Bayesian optimization). An example of active the selection of new experiments, and the execution and annota-
learning is the investigation of the effect of reaction conditions, tion of these experiments (Fig. 1). The active machine learning
such as temperature and pressure, on the conversion [10,11]. With query (phase 2) is determined through a so-called acquisition func-
Bayesian optimization, the goal is to find the optimal reaction con- tion, which is a measure of potential informativeness or optimality.
ditions to maximize this conversion [12–14]. In both cases, a The model needs the most informative subsequent data point,
design space is set up that defines the ranges of the studied vari- which is the point where the acquisition function is maximal for
ables by considering the objectives and the intrinsic limitations the selected query. The query is performed and new data is gath-
of the experimental tools. A machine learning model is then initial- ered (phase 3), after which the machine learning model is retrained
ized and trained using a small sample of labeled data, which comes (phase 1) and can now make improved predictions. This loop is
from experiments whose outcomes are known, stemming from lit- sequentially iterated until an optimum (Bayesian optimization) is
erature, previous experiments, or newly performed experiments. found or a sufficiently accurate model (active learning) is obtained.
In general, the amount of preliminary labeled data is very low. To further illustrate the workflow, we present the example of a
After initial training, the machine learning model is able to researcher examining the performance of a new catalyst for a
make rudimentary predictions in the design space. The model chemical process. The researcher either aims to investigate (with
can vaguely estimate where an optimum could be situated for active learning) or optimize (with Bayesian optimization) the effect
Bayesian optimization, or which experiment—the so-called of reaction variables (design space), such as the temperature,
query—is most informative for active learning. While the definition pressure, and reactant concentrations, on the desired product yield
and initialization of both active learning and Bayesian optimization (objective). First, initial experiments must be performed at a num-
are essentially the same (and are not even too different from a clas- ber of random combinations of temperature, pressure, and reactant
sic experimental campaign), the main differences and advantages concentrations. Next, the researcher initiates the active machine
are found in the model training. learning loop by training the machine learning model on these ran-
Active learning is purely based on exploration, to enable predic- domly picked experimental data points, after which the model pro-
tions of the design space that are as accurate as possible. Con- poses a new experiment. When using active learning, this
versely, Bayesian optimization balances both exploration and experiment is the most informative one; when optimizing with
exploitation in order to find the optimum in the design space, Bayesian optimization, this experiment is the most likely experi-
treating every iteration as the potentially final one. Exploitation ment to improve upon the desired product yield. The researcher
investigates areas with a high objective value to find an optimum performs the experiment and retrains the machine learning model,
nearby, whereas exploration discovers areas for which the predic- which now makes improved predictions. The experimental selec-
tions are unknown and therefore uncertain. Exploration requires a tion continues until the desired number of experiments is per-
measure of uncertainty in the predictions to identify which areas of formed and an optimal machine learning model or process
the design space remain unexplored [15]. Therefore, popular condition is obtained.
machine learning models for active machine learning are Gaussian
processes [16–19] and Bayesian neural networks [20–22], as these 1.2. Active machine learning in chemical engineering
allow an uncertainty estimation of their predictions. Another
advantage of Gaussian processes is that they deal very well with The applications of active machine learning span all the length
noisy measurements, which are inherent in real-life experiments. scales of chemical engineering, from ab initio calculations
By adding a noise term to the Gaussian process kernel, the machine [17,18,26] to material, molecule, and catalyst design [27–36], reac-
learning model can estimate the experimental uncertainty and tion design [12–14,37–42], and reactor design [43–45]. For exam-
allow optimal performance of the active machine learning method ple, the design of catalysts is an important asset in achieving
[16,23]. Neural networks can also be employed for active machine carbon neutrality, as catalysts can enable more sustainable pro-
learning purposes, but approximative methods such as Monte cesses and can increase the energy efficiency of chemical processes
Carlo dropout or model ensembling are required to estimate the in general [46]. However, catalyst design is still deemed an art
model uncertainty [11,24,25]. nowadays, as it mainly relies on high-throughput screening and
24
Y. Ureel, M.R. Dobbelaere, Y. Ouyang et al. Engineering 27 (2023) 23–30

limited theoretical relations, such as the Sabatier principle and lin-

ear scaling relations [47–50]. This makes catalyst design prone to
human bias, as researchers tend to exploit catalyst designs that
are known to work, which hampers real breakthroughs [51,52].
With active machine learning, this human bias is removed, and a
substantially larger fraction of the catalyst space can be studied.
Currently, the applications of active machine learning in catalysis
only consider a limited design space, varying only the catalyst
composition while maintaining the catalyst structure [53,54]. For
example, Zhong et al. [53] performed Bayesian optimization on
density functional theory (DFT) calculations to identify and synthe-
size promising electrocatalysts for the reduction of CO2, whereas
Nugraha et al. [54] determined the optimal composition of the
most active PtPdAu catalyst to electrocatalytically oxidize
methanol.
In reaction or process design, the goal of Bayesian optimization
is to determine the optimal operating conditions in order to max-
imize the product yields, minimize the emissions per product,
achieve the highest energy efficiency, and so forth. Optimization
of reaction conditions has been demonstrated multiple times,
including multi-objective reaction optimization with both discrete
and continuous variables, likely making this the most well-
developed field of active machine learning in chemical engineering
[12–14]. Shields et al. [39] applied Bayesian optimization to opti- Fig. 2. Three different types of thresholds for the breakthrough of active machine
mize the reaction conditions for a Mitsunobu reaction and learning (AML).
obtained an optimal yield (> 99%) for several non-intuitive reaction
conditions after 40 experiments, thereby overcoming the standard
machine learning and that an enormous experimental campaign
reaction yield of 60%. With active learning, the goal is to acquire
is required to make it feasible. Nugraha et al. [54] reported on an
reaction knowledge that can be used for reactor and catalyst
optimal catalyst composition performing only 47 of a total of
design, process control, or retrosynthesis. Eyke et al. [11] demon-
5151 possible experiments, as shown in Fig. 3. In their work,
strated the potential of active learning for DoE in reaction design
Bayesian optimization was employed to determine the optimal
by predicting reaction yields for combinations of catalysts and sol-
PtPdAu catalyst composition for the electrocatalytic oxidation of
vents with a minimum of available data. Recently, a DoE tool for
methanol. Similarly, Schweidtmann et al. [12] identified their
the study of chemical reactions was developed and validated on
Pareto front after 68 experiments for a four-dimensional reaction
the catalytic pyrolysis of plastic waste by Ureel et al. [10].
optimization. Moreover, Ureel et al. [10] showed that active learn-
CFD has become an important tool for reactors, optimization,
ing strategies are already beneficial for experimental campaigns
and trouble shooting. Bayesian optimization makes it possible to
consisting of as few as 18 experiments. These examples illustrate
find an optimal reactor configuration with a minimum of computa-
that both active learning and Bayesian optimization are already
tionally intensive CFD simulations. Park et al. [44] demonstrated
feasible for smaller datasets.
the power of multi-objective Bayesian optimization by maximizing
A second issue is related less to the experimental researcher
the gas holdup and minimizing the power consumption of a stirred
and more to the intrinsic algorithms. Initially, all active machine
tank reactor. Clearly integrating active machine learning in CFD
learning algorithms explore the entire design space, which can
allows for a faster and more efficient reactor design.
result in counterintuitive or trivial queries. Consequently, the
This survey shows that chemical engineering is a broad and
experimentalist loses confidence in the machine learning tool.
diverse research field with a whole spectrum of possible active
The initial selection of experiments does not rely on any prelimi-
machine learning applications. Nevertheless, the use of active
nary or physical knowledge within the machine learning models.
machine learning is not yet widespread, and there are some hur-
Therefore, this issue is related both to human bias and the percep-
dles to overcome before it can become a trusted asset in the chem-
tion of these algorithms by their users, and to the absence of pre-
ical engineer’s toolkit. In this perspective article, we focus on active
liminary knowledge within these models. Integrating process
machine learning as a DoE technique for an experimentalist and
knowledge beforehand in the machine learning model is the most
how to popularize it. We identify three types of thresholds: con-
powerful methodology to alleviate this problem. Such knowledge
vincing the experimental researcher, the flexibility of data creation,
can be incorporated via two different approaches: either through
and the robustness of active machine learning algorithms (Fig. 2).
the design of the machine learning model, such as a Gaussian pro-
In the following sections, we discuss each of these challenges
cess kernel [56], or through training on literature or simulation
and how they can be overcome.
data [57]. The incorporation of preliminary knowledge into active
machine learning models will be discussed in Section 4.1.
2. Convincing the researcher
2.2. Ease of use
2.1. Big data misconception
In active learning strategies, multiple factors are varied at the
At present, a knowledge gap exists between the experimentalist same time, whereas regular DoE strategies often vary a single fac-
community and machine learning experts [55]. This knowledge tor at a time. This makes the post-processing of the experiments
gap is the fundamental reason why active machine learning is less trivial, as the effects of the factors are not isolated. As a result,
not yet being systematically applied by experimentalists. First, a statistical analysis is required to draw conclusions from an exper-
there is a misconception that big data is mandatory for active imental campaign using an active learning strategy [58]. These
25
Y. Ureel, M.R. Dobbelaere, Y. Ouyang et al. Engineering 27 (2023) 23–30

Fig. 3. (a) Nugraha et al. [54] determined the optimal PtPdAu catalyst composition for the electrocatalytic oxidation of methanol by performing only 47 experiments, with a
higher peak current density denoting a better catalyst. (b) Contour plot of the effect of catalyst composition on peak current density, as determined by the 47 performed
experiments. Reproduced from Ref. [54] with permission.

tools are incorporated in regular DoE software but not in the active mesh independent [65]. When these constraints are non-trivial, a
machine learning packages that are currently available. This prob- separate machine learning model can be trained to learn the con-
lem is closely related to another issue that limits the applicability straints and enforce the viability of the simulations [43].
of active learning—namely, its ease of use. Many different active Another example with constrained experimental units is a high-
machine learning packages exist these days, such as Gryffin [59], throughput experimental campaign that is used to screen different
Phoenics [60], and BayesianOptimization [61] for Bayesian opti- catalytic materials. Within these units, several experimental vari-
mization, and Gaussian N-dimensional active learning framework ables, such as temperature and pressure, are often fixed for every
(GandALF) [10] or general and efficient active learning (GEAL) type of experiment per batch. This requires another constraint
[62] for active learning. However, most of the current active for the batch selection of these experiments, as the variables must
machine learning packages must be configured with Python, be fixed for all selected queries. To tune active machine learning
except for GandALF, which uses a csv spreadsheet. The use of these algorithms according to their application, a close collaboration
active machine learning tools requires programming skills, as they between the machine learning expert and the experimentalist is
offer no graphical user interface (GUI), hampering the usage of thus required. In this way, the benefits of applying active machine
these methodologies. Thus, at present, researchers that wish to learning are also available for less flexible experimental units.
use active machine learning must make a substantial time invest- Symbiosis between the experimentalist and the machine learn-
ment. This ‘‘activation barrier” is too high for many researchers, ing scientist will benefit both parties. First of all, it will extend the
particularly because of the required ability to code. fields of application for active machine learning as researchers
become more aware of the benefits of active machine learning. This
3. Improving the flexibility of data creation close collaboration will help in identifying useful features within
these active machine learning algorithms, such as blocking or auto-
3.1. Constrained active machine learning matic post-processing. More practical constraints might be added
to the experimental selection, such as the time or cost required
Active machine learning algorithms are often developed on sim- for a proposed experiment. Lastly, this collaboration between the
ulated data, where there are no practical limitations on the data experimentalist and the machine learning expert assists in inform-
creation side [32,36,63]. However, in real life, experimental units ing experimental researchers and removing the currently existing
or procedures do not allow this flexibility. For example, even a biases against active machine learning.
completely automated experimental unit often needs to heat up
or cool down, or requires time to stabilize, which slows down 3.2. Automation
the generation of a new data point when different temperatures
are selected by the algorithm. In addition, experiments are often In an ideal case, active machine learning is coupled with a flex-
performed in parallel (e.g., in high-throughput units), as opposed ible automated experimental unit or is even equipped by a robot
to the algorithms, which assume a sequential selection of experi- [12,14,66]. Thus, the control and optimization of the performance
ments. Therefore, active machine learning strategies should be of the experiments can become optimal, saving valuable time
constrained to the unit on which they are used, to allow for an and effort. Automated experimental units are increasingly being
optimal experimental efficiency that will make them applicable applied in molecular synthesis and chemical engineering, although
to real-world applications [64]. In the example above, it is often these units are not yet commonplace [67–69]. One requirement of
easier to heat an experimental unit than it is to cool it; therefore, automated robotic units is that they should be reconfigurable [70].
an extra constraint should be added to the algorithm to make it Moreover, they should have a broad application range and should
preferable to select experiments that increase rather than decrease not be limited to the investigation of a single reaction type or a
in temperature. narrow temperature range. Of course, the use of automated units
Next to constraints resulting from how the experimental equip- is not self-evident, as they are often expensive and are currently
ment operates, constraints can also be important for simulations not well-suited for every problem. For example, despite past efforts
[43,45]. Let us consider a case that involves optimizing a reactor [71], the automated synthesis and testing of catalysts is a challeng-
in silico using CFD. When defining the reactor geometry for CFD, ing task, especially when studying a broad design space [72]. By
it is not trivial that every type of geometry is feasible to simulate, coupling these systems with active machine learning techniques,
nor that the geometry can be properly meshed or the results are enormous time saving is expected for experimental campaigns,
26
Y. Ureel, M.R. Dobbelaere, Y. Ouyang et al. Engineering 27 (2023) 23–30

as this will speed up reaction and catalyst optimization, as well as strategies that can be employed to achieve this. When an active
the acquisition of scientific knowledge. A last threshold of these machine learning model is pretrained on literature data, an
automated units is the question of the safety of these units. By improved initial experimental selection is achieved that resolves
expanding the catalyst or reaction design space, safety concerns the issue of suboptimal initial selection that was mentioned earlier
increase, as doing so increases the probability that undesired reac- [57]. The incorporation of literature data is trivial when the exper-
tions will occur. Therefore, good chemical knowledge is still imental uncertainty is similar to that of the newly gathered data.
required when employing these units in order to identify and However, when the literature data is of better or inferior quality
incorporate safety constraints. Here, the definition of safety con- than the gathered data, it is important for the machine learning
straints again requires close collaboration between experimental model to be able to make a distinction between the two.
experts and machine learning scientists. Heteroscedastic machine learning models exist [63], but they do
not necessarily permit the incorporation of two separate noise fac-
tors, as the variation in noise is dependent on the variable in
4. Algorithm robustness heteroscedastic models. Conversely, multi-fidelity active machine
learning strategies make it possible to employ widely abundant
4.1. Data transfer low-quality data for accurate pretraining of the active machine
learning model [74–76]. These methods have been developed
When performing experiments, it is advantageous for the based on simulated ‘‘experimental” data only, but they are very
experiments to be widely applicable and to serve multiple pur- promising for improving the performance of active machine learn-
poses. The information gathered in experiments should be made ing tools when applied to real experimental data. Moreover, these
available according to the FAIR guiding principles (i.e., findability, multi-fidelity models can also be used for the incorporation of data
accessibility, interoperability, and reusability) and can then be of from a mechanistic model into the machine learning model. When
value for other researchers [73]. However, with active machine the uncertainty of the mechanistic model predictions is known, an
learning, a single objective is chosen, which determines the exper- appropriate distinction can be made between experimental data
imental selection. This hampers the applicability of the experi- and modeled data, both with their respective uncertainties, in
ments, as only one experimental output is well-studied. For the multi-fidelity model. In this way, additional mechanistic infor-
example, when investigating reactions, the conversion is typically mation can be incorporated into a machine learning model, which
selected as the output of interest; however, this limits the informa- improves the experimental selection.
tion on other properties, such as yields or selectivity. In the worst- Data that is closely related—but not similar in nature—can also
case scenario, the yields are not measured and no information is serve as an initialization for active machine learning models [77].
gathered; contrarily even when these yields are measured, it can- For example, when modeling reactions with one type of catalyst
not be guaranteed that all trends are considered in the example. As and literature data on another catalyst are available, this data
the goal of active machine learning is to model conversions, this might still contain valuable information for an active learning
method ignores the behavior of interesting reaction yields, which model [78]. With active transfer learning, the goal is to leverage
can result in trends remaining hidden. With Bayesian optimization, this knowledge from nearly similar data to obtain a machine learn-
this does not pose an issue, as the goal is to optimize an objective, ing model with an improved perception of the examined problem.
which makes the data per definition less generally applicable. Active transfer learning is the combination of the two main meth-
Multi-objective Bayesian optimization techniques exist, whereas ods of active machine learning and transfer learning to make
only single objective strategies are possible for active learning, machine learning less data intensive. With transfer learning,
meaning that all interesting outputs should be incorporated within (abundantly) available low-quality data is used to pretrain a
a single active learning objective [12,40,44]. Therefore, to ensure machine learning model, which is then refined with a limited
the reusability of the gathered data, it is important that not only amount of high-quality data. In this way, rudimentary physical
the modeled output but also other potential relevant outputs are knowledge is introduced into the machine learning model, which
measured during experiments. again improves the initial experimental selection. This methodol-
After creating data that is of wide interest, it is important to be ogy has been proven to work on the reaction yield classification
able to incorporate that knowledge into active machine learning of cross-coupling reactions, by pretraining a machine learning
tools. Fig. 4 summarizes the different data sources and modeling model on reactions with different nucleophiles [78].

Fig. 4. The incorporation of data from the literature, simulations, or expert knowledge into machine learning models via transfer learning or multi-fidelity models improves
the active machine learning performance.

27
Y. Ureel, M.R. Dobbelaere, Y. Ouyang et al. Engineering 27 (2023) 23–30

The reuse of literature data within active machine learning Aside from this intuitive approach, learned machine learning
applications will further enhance the performance of these tools. representations make it possible to create a continuous representa-
The first active transfer learning approaches are being developed tion, which ensures the validity of the proposed queries [80,81]. By
within chemical engineering, but further development of algo- training recently developed methodologies such as variational
rithms is crucial to make active transfer learning applicable within auto-encoders or generative adversarial neural networks on a set
all domains of chemical engineering. of synthesizable molecules or catalysts, a learned machine learning
representation—that is, a so-called latent space—can be developed,
4.2. Synthesizability ensuring the synthesizability of the proposed queries [80,82,83].
Upon this representation, additional constraints on the catalyst
Active machine learning can be used to determine the optimal or molecule can be enforced, according to the application [31].
query for either optimization or modeling purposes. However, for Finding an adequate representation is always important in
certain problems, it is not evident that these queries are exe- machine learning problems. For active machine learning, this rep-
cutable. For example, in catalyst or molecule design, novel com- resentation is essential in order to harmonize both synthesizability
pounds are proposed to synthesize and test the property of and creativity.
interest. Here, the representation of the catalyst or molecule is cru-
cial for the synthesizability of the queries. Synthesizability, which
is defined as the feasibility of the proposed queries, refers to 5. Conclusions and perspectives
whether the proposed catalysts or molecules can be synthesized,
as illustrated in Fig. 5. Often, a vector containing the catalyst com- Active machine learning is extremely well suited for use by
position is a simple representation of a catalyst [54,79]. This chemical engineering researchers to speed up experimental cam-
ensures the synthesizability of the catalyst but limits the design paigns ranging from molecule and catalyst design to reaction and
space explored by the active machine learning algorithm, as only reactor design. However, active machine learning is not well-
the composition is varied and no structural or geometrical proper- known among experimental researchers, and many active machine
ties are considered. Ideally, the complete catalyst space is consid- learning applications are not currently user friendly. Better collab-
ered for every problem by, for example, considering the complete oration between machine learning experts and chemical engineers
three-dimensional (3D) geometry as a representation of the cata- can overcome these barriers. Such interactions will also help to
lyst site or molecule. However, not every imaginable catalyst’s or tune active machine learning algorithms, depending on the applied
molecule’s 3D geometry can be synthesizable, so there is tradeoff (automated) experimental units and procedures, which will
between the magnitude of the design space, so-called creativity, improve the performance of these algorithms. A key barrier here
and synthesizability. is the suboptimal initial experimental selection, which can be over-
As illustrated by the previous example, the problem of synthe- come by integrating transfer learning and active learning with the
sizability essentially boils down to a problem of the machine learn- aid of multi-fidelity models. Moreover, the application domain of
ing representation upon which constraints are added to enforce active machine learning can be significantly extended by adapting
synthesizability. One intuitive approach is to use the synthesis pro- general active machine learning algorithms to obtain ‘‘tailor-made”
cess of the catalyst or molecule as the machine learning represen- algorithms, depending on the setup constraints. While the algo-
tation. A vector containing the catalyst composition, calcination rithms should be customized, the data should be generally usable,
temperature and time, and presence of ion exchange or impregna- such that performed experiments can serve multiple purposes. By
tion can be used to represent a catalyst. In this way, the synthesiz- harmonizing synthesizability and creativity, active machine learn-
ability of the queries is ensured, as every proposed recipe is ing is bound to make significant advances in the fields of molecule
executable. However, this representation does not necessarily and catalyst synthesis. Recent promising breakthroughs will allow
ensure an easy mapping to the property of interest, and an active machine learning to become an essential tool for the
increased amount of data might be required to model this relation. chemical engineer and will further facilitate autonomous and

Fig. 5. An illustration of synthesizability. A machine learning model proposes a query, which is essentially a vector representation of the catalyst. This query corresponds with
a catalyst, which can either be realistic and synthesizable (top) or unrealistic and non-synthesizable (bottom).

28
Y. Ureel, M.R. Dobbelaere, Y. Ouyang et al. Engineering 27 (2023) 23–30

efficient scientific discoveries, which will contribute to a more sus- [23] McHutchon A, Rasmussen C. Gaussian process training with input noise. In:
Shawe-Taylor J, Zemel R, Bartlett P, Pereira F, Weinberger KQ, editors.
tainable chemical industry in the future.
Proceedings of the 24th International Conference on Neural Information
Processing Systems; 2011 Dec 12–14; Granada, Spain; 2011. p. 1341–9.
[24] Zhang Y, Lee AA. Bayesian semi-supervised learning for uncertainty-calibrated
Acknowledgments
prediction of molecular properties and active learning. Chem Sci 2019;10
(35):8154–63.
Yannick Ureel, Maarten R. Dobbelaere, and Kevin De Ras respec- [25] Núñez M, Vlachos DG. Multiscale modeling combined with active learning for
tively acknowledge financial support from the Fund for Scientific microstructure optimization of bifunctional catalysts. Ind Eng Chem Res
2019;58(15):6146–54.
Research Flanders (FWO Flanders) through the doctoral fellowship [26] Sivaraman G, Krishnamoorthy AN, Baur M, Holm C, Stan M, Csányi G, et al.
grants (1185822N, 1S45522N, and 3F018119). The authors Machine-learned interatomic potentials by active learning: amorphous and
acknowledge funding from the European Research Council (ERC) liquid hafnium dioxide. NPJ Comput Mater 2020;6(1):104.
[27] Reker D, Schneider P, Schneider G, Brown JB. Active learning for computational
under the European Union’s Horizon 2020 research and innovation chemogenomics. Future Med Chem 2017;9(4):381–402.
programme (818607). [28] Brown KA, Brittman S, Maccaferri N, Jariwala D, Celano U. Machine learning in
nanoscience: big data at small scales. Nano Lett 2020;20(1):2–10.
[29] Hansen MH, Torres JAG, Jennings PC, Wang Z, Boes JR, Mamun OG, et al. An
Compliance with ethics guidelines atomistic machine learning package for surface science and catalysis. 2019.
arXiv:1904.00904.
[30] Griffiths RR, Hernández-Lobato JM. Constrained Bayesian optimization for
Yannick Ureel, Maarten R. Dobbelaere, Yi Ouyang, Kevin De Ras, automatic chemical design. 2017. arXiv:1709.05501.
Maarten K. Sabbe, Guy B. Marin, and Kevin M. Van Geem declare [31] Griffiths RR, Hernández-Lobato JM. Constrained Bayesian optimization for
automatic chemical design using variational autoencoders. Chem Sci 2020;11
that they have no conflict of interest or financial conflicts to
(2):577–86.
disclose. [32] Tran K, Ulissi ZW. Active learning across intermetallics to guide discovery of
electrocatalysts for CO2 reduction and H2 evolution. Nat Catal 2018;1
(9):696–703.
References [33] Kusne AG, Yu H, Wu C, Zhang H, Hattrick-Simpers J, DeCost B, et al. On-the-fly
closed-loop materials discovery via Bayesian active learning. Nat Commun
[1] Oxford Economics Ltd. The global chemical industry: catalyzing growth and 2020;11(1):5966.
addressing our world’s sustainability challenges. Oxford: Oxford Economics [34] Oftelie LB, Rajak P, Kalia RK, Nakano A, Sha F, Sun J, et al. Active
Ltd.; 2019. learning for accelerated design of layered materials. NPJ Comput Mater
[2] Lazić ŽR. Design of experiments in chemical engineering: a practical 2018;4(1):74.
guide. Weinheim: Wiley-VCH Verlag GmbH & Co. KGaA; 2006. [35] Kitchin JR. Machine learning in catalysis. Nat Catal 2018;1(4):230–2.
[3] Franceschini G, Macchietto S. Model-based design of experiments for [36] Jablonka KM, Jothiappan GM, Wang S, Smit B, Yoo B. Bias free multiobjective
parameter precision: state of the art. Chem Eng Sci 2008;63(19):4846–72. active learning for materials design and discovery. Nat Commun 2021;12
[4] Melnikov AA, Poulsen Nautrup H, Krenn M, Dunjko V, Tiersch M, Zeilinger A, (1):2312.
et al. Active learning machine learns to create new quantum experiments. Proc [37] Zhang C, Amar Y, Cao L, Lapkin AA. Solvent selection for Mitsunobu reaction
Natl Acad Sci USA 2018;115(6):1221–6. driven by an active learning surrogate model. Org Process Res Dev 2020;24
[5] Duong-Trung N, Born S, Kim JW, Schermeyer MT, Paulick K, Borisyak M, et al. (12):2864–73.
When bioprocess engineering meets machine learning: a survey from the [38] Clayton AD, Manson JA, Taylor CJ, Chamberlain TW, Taylor BA, Clemens G, et al.
perspective of automated bioprocess development. Biochem Eng J Algorithms for the self-optimisation of chemical reactions. React Chem Eng
2023;190:108764. 2019;4:1545–54.
[6] Olsson F. A literature survey of active machine learning in the context of [39] Shields BJ, Stevens J, Li J, Parasram M, Damani F, Alvarado JIM, et al. Bayesian
natural language processing. Kista: Swedish Institute of Computer Science; reaction optimization as a tool for chemical synthesis. Nature 2021;590
2009. (7844):89–96.
[7] Marin GB, Galvita VV, Yablonsky GS. Kinetics of chemical processes: from [40] Felton KC, Rittig JG, Lapkin AA. Summit: benchmarking machine learning
molecular to industrial scale. J Catal 2021;404:745–59. methods for reaction optimisation. Chem–Methods 2021;1(2):116–22.
[8] Settles B. Active learning. Cham: Springer Nature Switzerland AG; 2012. [41] Felton K, Wigh D, Lapkin A. Multi-task Bayesian optimization of chemical
[9] Frazier PI. A tutorial on Bayesian optimization. 2018. arXiv:1807.02811v1. reactions. 2020. ChemRxiv: 13250216.v1.
[10] Ureel Y, Dobbelaere MR, Akin O, Varghese RJ, Pernalete CG, Thybaut JW, et al. [42] Dogu O, Eschenbacher A, Varghese RJ, Dobbelaere M, D’Hooge DR, Van
Active learning-based exploration of the catalytic pyrolysis of plastic waste. Steenberge PHM, et al. Bayesian tuned kinetic Monte Carlo modeling of
Fuel 2022;328:125340. polystyrene pyrolysis: unraveling the pathways to its monomer, dimers, and
[11] Eyke NS, Green WH, Jensen KF. Iterative experimental design based on active trimers formation. Chem Eng J 2023;455:140708.
machine learning reduces the experimental burden associated with reaction [43] Tran A, Sun J, Furlan JM, Pagalthivarthi KV, Visintainer RJ, Wang Y. pBO-2GP-
screening. React Chem Eng 2020;5(10):1963–72. 3B: a batch parallel known/unknown constrained Bayesian optimization with
[12] Schweidtmann AM, Clayton AD, Holmes N, Bradford E, Bourne RA, Lapkin AA. feasibility classification and its applications in computational fluid dynamics.
Machine learning meets continuous flow chemistry: automated optimization Comput Methods Appl Mech Eng 2019;347:827–52.
towards the Pareto front of multiple objectives. Chem Eng J 2018;352:277–82. [44] Park S, Na J, Kim M, Lee JM. Multi-objective Bayesian optimization of chemical
[13] Amar Y, Schweidtmann AM, Deutsch P, Cao L, Lapkin A. Machine learning and reactor design using computational fluid dynamics. Comput Chem Eng
molecular descriptors enable rational solvent selection in asymmetric 2018;119:25–37.
catalysis. Chem Sci 2019;10(27):6697–706. [45] Morita Y, Rezaeiravesh S, Tabatabaei N, Vinuesa R, Fukagata K, Schlatter P.
[14] Clayton AD, Schweidtmann AM, Clemens G, Manson JA, Taylor CJ, Niño CG, Applying Bayesian optimization with Gaussian process regression to
et al. Automated self-optimisation of multi-step reaction and separation computational fluid dynamics problems. J Comput Phys 2022;449:110788.
processes using machine learning. Chem Eng J 2020;384:123340. [46] Friend CM, Xu B. Heterogeneous catalysis: a central science for a sustainable
[15] Thrun S. Exploration in active learning. In: Arbib MA, editor. The handbook of future. Acc Chem Res 2017;50(3):517–21.
brain theory and neural networks. Cambridge: MIT Press; 1995. p. 381–4. [47] Sabatier P. La catalyse en chimie organique. Paris: Hachette Livre; 1920.
[16] Rasmussen CE, Williams CKI. Gaussian processes for machine French.
learning. Cambridge: MIT Press; 2006. [48] Ichikawa S. Harmonious optimum conditions for heterogeneous catalytic
[17] Podryabinkin EV, Shapeev AV. Active learning of linearly parametrized reactions derived analytically with Polanyi relation and Bronsted relation. J
interatomic potentials. Comput Mater Sci 2017;140:171–80. Catal 2021;404:706–15.
[18] Vandermause J, Torrisi SB, Batzner S, Xie Y, Sun L, Kolpak AM, et al. On-the-fly [49] Landau RN, Korré SC, Neurock M, Klein MT, Quann RJ. Hydrocracking
active learning of interpretable Bayesian force fields for atomistic rare events. phenanthrene and 1-methyl naphthalene: development of linear free energy
NPJ Comput Mater 2020;6(1):20. relationships. In: Oballa M, editor. Catalytic hydroprocessing of petroleum and
[19] Riis C, Antunes F, Hüttel FB, Azevedo CL, Pereira FC. Bayesian active learning distillates. Boca Raton: CRC Press; 2020. p. 421–32.
with fully Bayesian Gaussian processes. 2022. arXiv:2205.10186. [50] Vijay S, Kastlunger G, Chan K, Nørskov JK. Limits to scaling relations between
[20] Blundell C, Cornebise J, Kavukcuoglu K, Wierstra D. Weight uncertainty in adsorption energies? J Chem Phys 2022;156(23):231102.
neural networks. In: Proceedings of the 32nd International Conference on [51] Hong X, Chan K, Tsai C, Nørskov JK. How doped MoS2 breaks transition-metal
Machine Learning; 2015 Jul 7–9; Lille, France; 2015. p. 1613–22. scaling relations for CO2 electrochemical reduction. ACS Catal 2016;6
[21] Gal Y, Islam R, Ghahramani Z. Deep Bayesian active learning with image data. (7):4428–37.
In: Proceedings of the 34th International Conference on Machine Learning; [52] Pérez-Ramírez J, López N. Strategies to break linear scaling relationships. Nat
2017 Aug 6–11; Sydney, NSW, Australia; 2017. p. 1183–92. Catal 2019;2(11):971–6.
[22] Hafner D, Tran D, Lillicrap T, Irpan A, Davidson J. Noise contrastive priors for [53] Zhong M, Tran K, Min Y, Wang C, Wang Z, Dinh CT, et al. Accelerated discovery
functional uncertainty. In: Proceedings of the 35th Uncertainty in Artificial of CO2 electrocatalysts using active machine learning. Nature 2020;581
Intelligence Conference; 2019 Jul 22–25; Tel Aviv, Israel; 2020. p. 905–14. (7807):178–83.

29
Y. Ureel, M.R. Dobbelaere, Y. Ouyang et al. Engineering 27 (2023) 23–30

[54] Nugraha AS, Lambard G, Na J, Hossain MSA, Asahi T, Chaikittisilp W, et al. [69] Mateos C, Nieves-Remacha MJ, Rincón JA. Automated platforms for reaction
Mesoporous trimetallic PtPdAu alloy films toward enhanced electrocatalytic self-optimization in flow. React Chem Eng 2019;4(9):1536–44.
activity in methanol oxidation: unexpected chemical compositions discovered [70] Eyke NS, Koscher BA, Jensen KF. Toward machine learning-enhanced high-
by Bayesian optimization. J Mater Chem A 2020;8(27):13532–40. throughput experimentation. Trends Chem 2021;3(2):120–32.
[55] Dobbelaere MR, Plehiers PP, Van de Vijver R, Stevens CV, Van Geem KM. [71] Hahndorf I, Buyevskaya O, Langpape M, Grubert G, Kolf S, Guillon E, et al.
Machine learning in chemical engineering: strengths, weaknesses, Experimental equipment for high-throughput synthesis and testing of
opportunities, and threats. Engineering 2021;7(9):1201–11. catalytic materials. Chem Eng J 2002;89(1–3):119–25.
[56] Duvenaud DK. Automatic model construction with Gaussian processes [72] Oh KH, Lee HK, Kang SW, Yang JI, Nam G, Lim T, et al. Automated synthesis and
[dissertation]. Cambridge: University of Cambridge; 2014. data accumulation for fast production of high-performance Ni nanocatalysts. J
[57] Wang Z, Dahl GE, Swersky K, Lee C, Mariet Z, Nado Z, et al. Pre-training helps Ind Eng Chem 2022;106:449–59.
Bayesian optimization too. 2022. arXiv:220703084. [73] Wilkinson MD, Dumontier M, Aalbersberg IJ, Appleton G, Axton M, Baak A,
[58] Symoens SH, Aravindakshan SU, Vermeire FH, De Ras K, Djokic MR, Marin GB, et al. The FAIR Guiding Principles for scientific data management and
et al. QUANTIS: data quality assessment tool by clustering analysis. Int J Chem stewardship. Sci Data 2016;3(1):160018.
Kinet 2019;51(11):872–85. [74] Greenman KP, Green WH, Gómez-Bombarelli R. Multi-fidelity prediction of
[59] Häse F, Aldeghi M, Hickman RJ, Roch LM, Aspuru-Guzik A. Gryffin: an molecular optical peaks with deep learning. Chem Sci 2022;13(4):1152–62.
algorithm for Bayesian optimization of categorical variables informed by [75] Pilania G, Gubernatis JE, Lookman T. Multi-fidelity machine learning models
expert knowledge. Appl Phys Rev 2021;8(3):031406. for accurate bandgap predictions of solids. Comput Mater Sci
[60] Häse F, Roch LM, Kreisbeck C, Aspuru-Guzik A. Phoenics: a Bayesian optimizer 2017;129:156–63.
for chemistry. ACS Cent Sci 2018;4(9):1134–45. [76] Folch JP, Lee RM, Shafei B, Walz D, Tsay C, van der Wilk M, et al. Combining
[61] Snoek J, Larochelle H, Adams RP. Practical Bayesian optimization of machine multi-fidelity modelling and asynchronous batch Bayesian optimization.
learning algorithms. In: Pereira F, Burges CJ, Bottou L, Weinberger KQ, editors. Comput Chem Eng 2023;172:108194.
Proceedings of the 25th International Conference on Neural Information [77] Mao S, Wang B, Tang Y, Qian F. Opportunities and challenges of artificial
Processing Systems; 2012 Dec 3–6; Lake Tahoe, NV, USA. Red Hook: Curran intelligence for green manufacturing in the process industry. Engineering
Associates Inc.; 2012. p. 2951–9. 2019;5(6):995–1002.
[62] Xie Y, Tomizuka M, Zhan W. Towards general and efficient active learning. [78] Shim E, Kammeraad JA, Xu Z, Tewari A, Cernak T, Zimmerman PM. Predicting
2021. arXiv:211207963. reaction conditions from limited data through active transfer learning. Chem
[63] Griffiths RR, Aldrick AA, Garcia-Ortegon M, Lalchand V, Lee AA. Achieving Sci 2022;13(22):6655–68.
robustness to aleatoric uncertainty with heteroscedastic Bayesian [79] Kim M, Ha MY, Jung WB, Yoon J, Shin E, Kim ID, et al. Searching for an optimal
optimisation. Mach Learn Sci Technol 2021;3(1):015004. multi-metallic alloy catalyst by active learning combined with experiments.
[64] Hickman RJ, Aldeghi M, Häse F, Aspuru-Guzik A. Bayesian optimization with Adv Mater 2022;34(19):2108900.
known experimental and design constraints for chemistry applications. Digit [80] Gómez-Bombarelli R, Wei JN, Duvenaud D, Hernández-Lobato JM,
Discov 2022;1:732–44. Sánchez-Lengeling B, Sheberla D, et al. Automatic chemical design using a
[65] Habashi WG, Dompierre J, Bourgault Y, Ait-Ali-Yahia D, Fortin M, Vallet MG. data-driven continuous representation of molecules. ACS Cent Sci 2018;
Anisotropic mesh adaptation: towards user-independent, mesh-independent 4(2):268–76.
and solver-independent CFD. Part I: general principles. Int J Numer Meth [81] Shang C, You F. Data analytics and machine learning for smart process
Fluids 2000;32(6):725–44. manufacturing: recent advances and perspectives in the big data era.
[66] Burger B, Maffettone PM, Gusev VV, Aitchison CM, Bai Y, Wang X, et al. A Engineering 2019;5(6):1010–6.
mobile robotic chemist. Nature 2020;583(7815):237–41. [82] Sanchez-Lengeling B, Outeiral C, Guimaraes GL, Aspuru-Guzik A. Optimizing
[67] Hoffer L, Voitovich YV, Raux B, Carrasco K, Muller C, Fedorov AY, et al. distributions over molecular space. An objective-reinforced generative
Integrated strategy for lead optimization based on fragment growing: the adversarial network for inverse-design chemistry (ORGANIC). 2017.
diversity-oriented-target-focused-synthesis approach. J Med Chem 2018;61 ChemRxiv: 5309668.v3.
(13):5719–32. [83] Jensen Z, Kwon S, Schwalbe-Koda D, Paris C, Gómez-Bombarelli R, Román-
[68] Bédard AC, Adamo A, Aroh KC, Russell MG, Bedermann AA, Torosian J, et al. Leshkov Y, et al. Discovering relationships between OSDAs and zeolites
Reconfigurable system for automated optimization of diverse chemical through data mining and generative neural networks. ACS Cent Sci 2021;7
reactions. Science 2018;361(6408):1220–5. (5):858–67.

30