An experimental architecture for AI-guided inverse design of materials and material-enabled systems.
The project is intentionally starting as a design repository, not as a software implementation. Its purpose is to build a reproducible search-and-validation system in which AI chooses what to evaluate next while numerical physics remains the source of quantitative evidence.
The first concrete research track is atmospheric-entry working-fluid discovery.
The goal is not to design a water-cooled heat shield.
The goal is to discover or formulate a liquid / liquid mixture that minimizes the working-fluid mass required to:
- keep the vehicle within thermal limits;
- provide any useful additional braking through directed vapor discharge;
- remain physically storable, deliverable and compatible with the system.
Water is used only as a reference and evaluator-validation fluid because its properties are well characterized.
Read Project Goal first.
mission + architecture
|
v
candidate fluid / mixture
|
v
known or predicted properties
|
v
fluid evaluator
|
+---- thermal contribution
|
+---- vapor / nozzle contribution
|
+---- aerodynamic interaction
|
v
required working-fluid mass
|
v
compare with water reference
|
v
AI / optimizer chooses next candidate
|
+--------------------> repeat
The same evaluator should accept water, known liquids, mixtures and later hypothetical molecular candidates without changing the project objective.
Water is a calibration point, not the desired end product.
water
-> verify property model
-> verify conservation
-> obtain reference score
candidate X
-> same evaluator
-> compare required mass with water
A scientifically meaningful result is of the form:
Under the stated mission, hardware assumptions and model fidelity, candidate X requires Y% less working-fluid mass than the water reference while satisfying the same constraints.
The architecture remains general enough to support other domains, including crystal discovery.
Autonomous Materials Lab
|
+-- fluid / mixture inverse design
| -> thermodynamics
| -> MD / chemistry
| -> aerothermal system evaluator
|
+-- crystal discovery
-> structure generation
-> ML screening
-> DFT
-> phonons / stability
These tracks share:
- orchestration;
- provenance;
- budget accounting;
- AI experiment selection;
- uncertainty handling;
- high-fidelity escalation.
They do not share the same physical evaluator.
The first prototype is designed around:
- 2 × NVIDIA RTX 4090;
- multicore x86_64 CPU;
- local NVMe storage;
- optional university/HPC access.
The first reduced fluid evaluator may run almost entirely on CPU. The GPUs become important for high-throughput ML, molecular simulation, surrogate models and later expensive candidate screening.
Domain-dependent candidates include:
- NumPy / SciPy — reduced-order models and optimization;
- validated thermophysical-property backend;
- Cantera — thermodynamics / reacting-gas chemistry where applicable;
- LAMMPS — molecular dynamics;
- CP2K — selected ab-initio MD / molecular reference calculations;
- SU2 or another open CFD solver — later aerothermal interaction modelling;
- custom trajectory / thermal / injection evaluator.
- ASE;
- pymatgen;
- MatterGen or another generator;
- MACE / CHGNet;
- Quantum ESPRESSO;
- CP2K;
- Phonopy.
- Python 3.12 production baseline;
- SQLite/PostgreSQL + artifact storage;
- Pydantic or equivalent typed schemas;
- LLM / reasoning model as scientific controller;
- optional AiiDA when remote-workflow complexity justifies it.
- The search target must be explicit.
- Water is a benchmark, not the design objective.
- Physics outranks language-model preference.
- Every rejected candidate has an explicit reason.
- A numerical failure is not a physical rejection.
- Expensive calculations are treated as budgeted resources.
- Every result is reproducible from stored inputs, versions and settings.
- The controller must explain why a calculation was selected.
- Candidate comparison must use the same evaluator and comparable system assumptions.
- Project goal
- Architecture
- Liquid TPS / working-fluid use case
- Generic fluid evaluator contract
- Water reference benchmark
- MVP plan
- Discovery loop
- AI controller
- Candidate lifecycle
- Compute strategy
- Development and compute environment
- Source development and optimization
- Preflight checklist
Start here:
- Ambient-liquid heat-sink screening
- Cryogenic / storage-state enthalpy screening
- Reactive fluid + air equilibrium check
- Hydrogen / air ignition-delay sweep
- Low-fidelity end-to-end Earth entry evaluator
- Structure-derived property predictor V5b
- Constrained inverse molecular search V5c
- High-throughput molecular search V5d
- Active-learning calibration expansion V5e
The first experiments isolate thermodynamics and chemical kinetics. The entry evaluator is the first system-level calculation: it propagates a configurable Earth-entry trajectory, evaluates atmosphere, stagnation-point convective and radiative heating, wall thermal response, coolant enthalpy and required coolant mass flow. Independent cases can be distributed across CPU processes. It remains intentionally below CFD / DSMC fidelity and does not yet model detailed boundary-layer blowing, porous flow or nozzle geometry.
Phase 1 — reduced-order end-to-end entry evaluator and working-fluid comparison.
The repository now contains local property / chemistry screens, a trajectory-level evaluator, and a structure-derived property-prediction track. The staged entry work includes V1 numerical verification, V2 physical-sensitivity screening, V3 cross-model Earth aeroheating comparison, V4 ignition-delay constrained hydrogen cooling, and V5a known binary liquid-mixture modeling. V5b adds SMILES -> GC-PC-SAFT/Joback -> PropertyProvider -> entry evaluator; V5b-2 adds a broad blind holdout and structural applicability-domain uncertainty; V5b-2.1 adds failure taxonomy, repeated-split robustness and leave-one-family-out challenges. V5c-1 then performs deterministic constrained molecular generation and uncertainty-aware ranking against Water. V5c-2 expands the parameter-supported molecular space, adds a thermodynamic property pre-screen, and separates statistically rankable candidates from exploratory new functional families. V5d-1 replaces hand-written enumeration with deterministic graph mutation, evolutionary beam search, diversity preservation, and a hierarchical high-throughput funnel targeting up to 120k new structures. V5e-1 closes the loop with active-learning acquisition: it selects a diverse, property-relevant set of exploratory molecules for independent reference-data resolution and later V5b recalibration; selection alone does not expand the calibrated domain. V5e-2 automates the first reference-resolution handoff using exact PubChem identity checks, exact CoolProp identifier matching, NIST WebBook corroboration, source caching and an explicit calibration-quality gate. V5e-2.1 extends identity-resolved targets beyond CoolProp with an explicit multi-source empirical allowlist over chemicals/thermo, preserves rejected predicted methods in an audit trail, and emits property-calibration anchors for V5b-3 without mislabeling them as entry-level ground truth. None of these reduced-order stages establishes flight-level physical validity.