How it works
LeilMat under the hood
A short, honest tour of the tool: what each surface actually does, which open-source projects make it possible, and where the numbers come from. Built for working metallurgists who want to know what they’re trusting.
00 · Quickstart
From element to alloy in four steps
- 1.Set a composition. Drag the element sliders in the sidebar or pick from the periodic table — the total auto-normalises to 100 %.
- 2.Predict properties. One click gives mechanical + physical properties with calibrated uncertainty — and unlocks the ten process simulations.
- 3.Simulate & optimise. Run CALPHAD, Scheil, KWN, grain structure or CPFEM; sweep a variable; or set targets and let inverse design propose candidates. The AI copilot suggests next moves you apply with one tap.
- 4.Export. The whole session becomes a multi-page, audit-grade PDF with provenance on every number.
Every result carries a calibration label (deterministic / quantitative / literature-correlation / qualitative) telling you how far to trust its absolute numbers. These are screening tools — validate against experiment before production.
01 · What it does
An AI-native canvas for alloy design
LeilMat sits in the space between a CALPHAD tool (Thermo-Calc, JMatPro) and an ML platform (Citrine). The canvas is one spatial editor where you drag composition sliders, pick a temper, and see live property predictions with calibrated uncertainty. From there you can run physical simulations (CALPHAD equilibrium, Scheil solidification, KWN precipitation, 1-D Fick diffusion, anisotropic phase-field dendrites) without leaving the page.
An agentic AI copilot reads the full session — composition, goals, constraints, what you’ve explored — and emits clickable action pills rather than just narrating advice. “Set Mg to 4 %, cap density at 2.7, run Predict” — three taps, three slider mutations, one prediction. The whole design exports as a multi-page PDF with provenance footnotes on every number.
02 · Architecture
Three layers, one canvas
The fast statistical layer (top) gives instant feedback while you slide. The physics layer (middle) runs the deterministic simulators the predictions are validated against. The data layer (bottom) is what makes any of it credible — bundled CALPHAD thermodynamic databases, the Materials Project DFT corpus, handbook-grade experimental measurements with full citations.
03 · Open-source stack
What's underneath
LeilMat is built on a stack of mature open-source projects. Where a paid tool was an option we picked the open one, partly for cost but mostly because users in regulated industries need to inspect what they’re running.
Scientific simulation
- pycalphad— CALPHAD equilibrium + Scheil-Gulliver solidification. BSD-licensed; bundled with 7 TDB databases (COST507, mc_fecocrnbti, alcrni, alni, alzn, alcocrni, Fe-C binaries).
- pymatgen— Materials genomics — structure handling, Materials Project I/O, composition arithmetic.
- matminer— Featurisers for ML on materials (Magpie, composition descriptors).
- mp-api— Materials Project REST client — DFT-computed properties for ~150k inorganic compounds.
- Taichi— JIT-compiled GPU-friendly phase-field solver. Drives the anisotropic KWC dendrite simulator (~3.6 min for a 6-arm dendrite at 400² on CPU).
- FiPy— Finite-volume PDE solver used in the simpler phase-field path.
- NumPy / SciPy— Crank-Nicolson tridiagonal solver for the 1-D Fick diffusion mini-sim; numerical backbone throughout.
ML & optimization
- scikit-learn— Random Forest models per alloy family (Al / Ti / Ni / Steel). Temper-aware one-hot + strength-index features.
- BoTorch— Multi-objective Bayesian optimisation for inverse design. Conformal-CI-aware acquisition function.
- PyTorch— ML framework under BoTorch and MatGL.
- MatGL / M3GNet— Graph neural network for formation-energy predictions; ML proxy for DFT.
04 · Data sources
Where the numbers come from
Thermodynamic databases (bundled, BSD)
- COST507— Light-metal binary + ternary assessments. Reference for Al-Mg, Al-Si, Al-Cu, Al-Zn.
- mc_fecocrnbti— Multi-component Fe-Co-Cr-Ni-Nb-Ti — covers most stainless + superalloy ranges.
- alcrni / alni / alzn / alcocrni— Smaller targeted assessments for specific Al-X systems.
Experimental measurements
- LeilMat alloy dataset— ~150 entries across Al / Ti / Ni / Steel with composition + temper + yield strength + density + hardness. Every row has a handbook citation (ASM, MIL-HDBK-5J, Boyer, Donachie, Reed). Defence-in-depth holdout filter prevents the model from learning a row it's about to be tested on.
- Smithells Metals Reference Book (7e)— Source for diffusivity D₀ / Q in the 1-D Fick mini-sim (Al-Mg, Al-Cu, Al-Zn, Fe-Mn, Fe-C, Ni-Cr, Ni-Al pairs).
- Shewmon — Diffusion in Solids— Cross-reference + auxiliary D-coefficients.
External APIs
- Materials Project— DFT-computed formation energies, convex-hull stability, ~150k inorganic structures. Accessed via mp-api with the user’s own MP_API_KEY.
- NIMS MatNavi (planned)— Free Japanese national materials database — adding ~3k entries via responsible scraping is on the data-ingest roadmap.
Models we trained
- Random Forest yield strength (per family)— scikit-learn RF. Features: ~50 composition descriptors + temper one-hot + strength index. CV MAE: Al 41 MPa, Ti 101 MPa, Ni 73 MPa, Steel 92 MPa. Conformal calibration on held-out folds gives 95 % CIs.
- OOD detector— Per-family Mahalanobis-distance gate. Compositions far from training data get an out-of-domain flag in the UI so the user knows the prediction is extrapolative.
- Hierarchical cost model— 49-test scientific harness; 17/17 quoted alloys within factor-of-2 of supplier prices.
05 · Scientific references
What the math is based on
- Kampmann & Wagner (1984) → Robson (2004) — numerical precipitation kinetics. Drives the KWN solver and the aging-playback scrubber.
- Scheil-Gulliver model — non-equilibrium solidification microsegregation. Implemented on top of pycalphad.
- Crank (1975) — The Mathematics of Diffusion — Crank-Nicolson finite-difference scheme used by the 1-D Fick mini-sim. Unconditionally stable + second-order accurate.
- Hall-Petch + Hollomon — yield-strength + work-hardening contributions in the physics-aware feature set.
- Voigt-Reuss-Hill averaging — elastic moduli; gives the YOUNGS_MODULUS card its VRH-branded value.
- Karma & Rappel (1998) anisotropic phase-field — the KWC dendrite simulator’s formulation.
- Smithells Metals Reference Book (7e), 1992 — diffusivity tables + interfacial energies + reference alloys.
- Shewmon, P. G. (1989) Diffusion in Solids — auxiliary D coefficients + theoretical framing.
06 · For developers
Integration
LeilMat is currently in private beta. Integration access and a full API specification are available on request through the Leil team. All cost-bearing endpoints require a Supabase JWT in the Authorization header; rate limits, daily-budget caps, and per-row security policies apply at the same layer the canvas uses.