SCALE: Simulation-Calibrated Amortized Learning for Energy Materials (A hybrid architecture connecting deterministic modeling, real-world data, and transformer-scale inference for accelerated energy-materials discovery)

Kuan Huang; Bo Bai

cs.LGSep 3, 2026

SCALE: Simulation-Calibrated Amortized Learning for Energy Materials (A hybrid architecture connecting deterministic modeling, real-world data, and transformer-scale inference for accelerated energy-materials discovery)

Authors: Kuan Huang, Bo Bai

Organizations: Alpha Ladder Particle Pte. Ltd., 168 Robinson Road, #26-02, Singapore 068912

Abstract

Energy systems face converging pressures for security, affordability, resilience, and sustainability, creating a need for faster discovery of deployable energy materials. Here we introduce SCALE (Simulation-Calibrated Amortized Learning for Energy Materials), a physics-grounded, real-world-data-calibrated learning architecture that connects deterministic scientific operators, experimental calibration, expanded calibrated label generation, and transformer-scale inference. SCALE converts selected high-cost mechanistic computation and measured evidence into reusable models for rapid screening, ranking, inverse design, and active learning. We formulate the framework, identify ten method-based application regimes, and demonstrate SCALE for solid-state metal-hydride hydrogen-storage capacity prediction. In this implementation, a hydride phase-equilibrium capacity operator is calibrated against 381 measured ML-HydPARK capacity anchors and used to generate 5,000 candidate-condition-prototype teacher labels. A crystallographically anchored periodic-graph representation preserves atomic sites, periodic neighbor relationships, and local metal environments absent from formula-only encodings. An edge-biased graph transformer with 2.90 million parameters reproduces calibrated teacher labels with five-fold surrogate fidelity of MAE 0.0582 wt% H2, RMSE 0.0833 wt% H2, R2 = 0.9927, and Pearson r = 0.9963. Post hoc attention analysis suggests that SCALE learns chemically organized element groupings and metal-metal relationships consistent with established hydride chemistry, without chemistry-group labels as supervision. Once trained, SCALE shifts million-candidate evaluation from repeated deterministic workflow execution to batched learned inference, reducing per-candidate screening cost by approximately 10^7-10^8 while retaining links to simulation and experimental evidence.

Figures & tables

Panel of Figure 1 — Figure 1: SCALE separates calibrated evidence generation from deployment inference. Deterministic operators and experiments are used offline to construct calibrated teacher labels for training; after training, large candidate permutations are evaluated by the learned model, and active-learning outputs return to experimental validation.

Category	Representative deterministic tools/workflows	Representative energy materials	Representative studies and SCALE role	Deterministic workload
Electronic-structure screening	VASP, Quantum ESPRESSO, CP2K, CASTEP, ABINIT, GPAW, atomate, AiiDA	Battery cathodes/anodes, solid electrolytes, perovskite PV absorbers, electrocatalysts, thermoelectrics, permanent magnets, hydrides, corrosion alloys, fusion alloys	Materials Project, AFLOW, and OQMD demonstrate high-throughput first-principles screening [ 9 , 10 , 11 ] ; atomate automates first-principles workflows [ 12 ] . SCALE can calibrate electronic-structure outputs against measured performance and distill them into composition/structure transformers.	Hours–days per structure; $10^{4}$ – $10^{7}$ core-hours per campaign.
Finite-temperature atomistic transport	AIMD, NEB, CP2K, VASP-AIMD, Quantum ESPRESSO CP, VTST	Solid electrolytes, hydride diffusion pathways, perovskite ion migration, catalytic surfaces, radiation-defect migration	First-principles dynamics and migration-barrier workflows provide mechanistic transport labels. SCALE can combine these labels with measured conductivity, diffusion, or degradation data for rapid transport screening.	Hours–days per path; $10^{3}$ – $10^{6}$ core-hours per campaign.
Classical and reactive molecular dynamics	LAMMPS, GROMACS, OpenMM, AMBER, CHARMM, NAMD, ReaxFF	Polymer electrolytes, PEM/AEM membranes, MOFs, CO 2 sorbents, SEI/CEI layers, catalyst-ionomer films, phase-change materials, thermal-interface materials	Molecular simulation has enabled large adsorption and transport studies in porous materials [ 31 ] . SCALE can convert hydration, uptake, transport, swelling, morphology, and degradation simulations into calibrated transformer labels.	Hours–days per system; $10^{3}$ – $10^{6}$ core-hours per campaign.
Machine-learned interatomic potentials	M3GNet, CHGNet, MatterSim-style potentials, DeePMD-kit, MACE, NequIP, GAP	Batteries, hydrides, high-entropy alloys, catalysts, perovskites, MOFs, fusion alloys, thermoelectrics	M3GNet and CHGNet show broad learned-potential transfer [ 16 , 17 ] ; large atomistic models further broaden this direction [ 18 ] . SCALE can place experimental calibration above ML-potential simulations.	High training cost; low–moderate inference cost per trajectory.
Thermodynamics and phase equilibria	Thermo-Calc, Pandat, FactSage, OpenCalphad, pycalphad	Hydrides, high-entropy alloys, molten salts, battery alloys, fusion blankets, nuclear alloys, thermoelectrics, phase-change alloys	pycalphad provides open CALPHAD infrastructure [ 32 ] . SCALE can learn calibrated phase-stability, phase-fraction, and plateau-pressure response surfaces.	Seconds–minutes per query after database assessment.
Reaction kinetics and kinetic Monte Carlo	Cantera, CHEMKIN, COPASI, Zacros, custom KMC solvers	Polymerization, electrolyte aging, CO 2 reduction, OER/HER/ORR catalysts, ammonia cracking, liquid organic hydrogen carriers	Zacros supports lattice KMC for catalytic surfaces [ 33 ] . SCALE can convert mechanism-based simulations and reaction experiments into fast reaction-outcome models.	Seconds–hours per ODE case; $10^{2}$ – $10^{5}$ core-hours for KMC.

Table 1: Method-based benchmark regimes for SCALE. Compute-density values are order-of-magnitude planning estimates if only deterministic models are used directly for screening before transformer amortization; actual cost depends on system size, convergence criteria, hardware, licensing, and campaign design.

Step	Stage	Input	Output	Tools and method	Computational burden
1	Parent-alloy definition	Formula, material class, $T$ , and $P_{\mathrm{H_{2}}}$	Parsed composition, molar mass, and condition record	pymatgen and validated composition parsing [ 45 ]	Low; linear in the number of elements, $O(E)$ .
2	Parent-structure resolution	Parsed formula and parent-only structure sources	Exact parent CIF or ordered parent approximant	Materials Project or OQMD retrieval; deterministic supercell and sublattice substitution [ 9 , 11 , 45 ]	Low–moderate; API/cache dominated for exact structures, increasing with approximant size.
3	Parent quantum relaxation	Parent CIF, pseudopotentials, $k$ -mesh, cutoffs, and smearing	Relaxed or plateau-accepted parent geometry and energy	Quantum ESPRESSO variable-cell relaxation with PBE pseudopotentials [ 46 ]	High; approximately $C_{\mathrm{vc}}(N,K,I)\sim N^{2.5\text{--}3}KI_{\mathrm{scf}}I_{\mathrm{geo}}$ .
4	Parent-structure analysis	Relaxed parent geometry	Distances, symmetry/prototype, and candidate interstitial regions	pymatgen , spglib , periodic distance analysis, and Voronoi/grid void search [ 45 , 47 ]	Low–moderate; pair analysis $O(N^{2})$ , with void-search cost set by grid density.
5	Hydrogen-loading hypotheses	Candidate interstices, formula units, and collision constraints	Parent+H structures across discrete loading states	Deterministic H-site placement, loading enumeration, and structure/input writers	Moderate; scales with candidate sites and loading branches, approximately $O(BN)$ .
6	Fast quantum screening	Parent+H candidates and parent/H 2 references	Ranked loading candidates	Fixed-cell or short-relaxation Quantum ESPRESSO screening, energy parsing, and formation-energy ranking	High, but screening grade; approximately $B\,C_{\mathrm{scf}}(N_{H},K,I)$ .

Table 2: Hydride phase-equilibrium capacity operator used in the metal-hydride implementation. The workflow begins with the hydrogen-free parent alloy, enumerates hydrogen-loading hypotheses without using measured capacity or known hydride stoichiometry, and promotes selected branches to full first-principles confirmation. Computational-burden expressions describe deterministic work rather than hardware-specific wall time. Because the present implementation evaluates the SCALE architecture rather than establishing a definitive metal-hydride phase diagram, Step 7 was benchmarked on representative candidate branches to quantify the computational burden of terminal variable-cell confirmation but was not executed exhaustively across the expanded corpus. Note:

N

is the number of atoms in the quantum cell,

K

is the number of irreducible

k

-points,

I_{\mathrm{scf}}

is the number of electronic iterations per geometry step,

I_{\mathrm{geo}}

is the number of ionic or cell-relaxation steps,

B

is the hydrogen-loading ladder size, and

m

is the number of branches promoted to full relaxation. Heavy elements, semicore states, metallic smearing, low symmetry, and large ordered approximants can substantially increase the computational prefactor.

Model	Validation split	$n$	MAE	RMSE	$R^{2}$	Pearson $r$
Hydride capacity operator	all rows	381	0.623	0.788	0.274	0.823
Ridge calibration	random 5-fold	381	0.291	0.450	0.763	0.875
Histogram-gradient boosting	random 5-fold	381	0.230	0.371	0.839	0.916
Random forest	random 5-fold	381	0.234	0.367	0.843	0.918
Extremely randomized trees	random 5-fold	381	0.219	0.380	0.831	0.912
Extremely randomized trees	material-class grouped	381	0.405	0.610	0.565	0.761

Table 3: Experimental calibration accuracy against ML-HydPARK measured H 2 wt%. The selected calibration uses deployable chemistry/condition descriptors and deterministic-operator outputs; raw operator performance directly compares uncalibrated simulation-derived capacity to experiment.

Split	$n$	MAE	RMSE	$R^{2}$	Pearson $r$	Bias
Overall OOF	5000	0.0582	0.0833	0.9927	0.9963	$+0.0022$
Fold 1	1000	0.0593	0.0901	0.9919	0.9960	$+0.0080$
Fold 2	1000	0.0591	0.0849	0.9929	0.9964	$-0.0017$
Fold 3	1000	0.0582	0.0825	0.9920	0.9960	$+0.0021$
Fold 4	1000	0.0589	0.0810	0.9932	0.9966	$-0.0026$
Fold 5	1000	0.0556	0.0773	0.9934	0.9967	$+0.0054$

Table 4: Overall and fold-level graph-transformer surrogate fidelity against calibrated metal-hydride H 2 wt% teacher labels. The overall row reports metrics calculated from all 5,000 out-of-fold predictions; fold rows demonstrate consistency across independently trained models.

Materials Discovery

See all 29 papers

Mar 13, 2026

Foundation-Model Surrogates Enable Data-Efficient Active Learning for Materials Discovery

Active learning (AL) has emerged as a powerful paradigm for accelerating materials discovery by iteratively steering experiments toward promising candidates, reducing the number of costly synthesis-and-characterization cycles needed to identify optimal materials. However, current AL relies predominantly on Gaussian Process (GP) and Random Forest (RF) surrogates, which suffer from complementary limitations: GP underfits complex composition-property landscapes due to rigid kernel assumptions, while RF produces unreliable heuristic uncertainty estimates in small-data regimes. This small-data challenge is pervasive in materials science, making reliable surrogate modeling extremely difficult with models trained from scratch on each new dataset. Here we propose In-Context Active Learning (ICAL), which addresses this bottleneck by replacing conventional surrogates with TabPFN, a transformer-based foundation model (FM) pre-trained on millions of synthetic regression tasks to meta-learn a universal prior over tabular data, upon which TabPFN performs principled Bayesian inference in a single forward pass without dataset-specific retraining, delivering strong small-data regression performance and well-calibrated predictive uncertainty (required for effective AL). We benchmark ICAL against GP and RF across 10 materials datasets and TabPFN wins on 8 out of 10 datasets, achieving a mean saving of 52% in extra evaluations relative to GP and 29.77% relative to RF. Cross-validation analysis confirms that TabPFN's advantage stems from superior uncertainty calibration, achieving the lowest Negative Log-Likelihood and Area Under the Sparsification Error curve among all surrogates. These results demonstrate that pre-trained FMs can serve as effective surrogates for active learning, enabling data-efficient discovery across diverse materials systems and small-data experimental sciences.

Jeffrey Hu, Rongzhi Dong, Ying Feng +2

Department of Materials Science and Engineering, University of Illinois Urbana Champaign, Champaign,IL, USA · Department of Computer Science & Engineering, University of South Carolina, Columbia, SC, USA · Zuoyue Honors College, Hangzhou Dianzhi University, Hangzhou, China +1

Active Learning Data-Efficient Learning Surrogate Modeling Materials Property Prediction Tabular Foundation Models

Ask PDF

May 15, 2026

PRISMat: Policy-Driven, Permutation-Invariant Autoregressive Material Generation

Rapid identification of candidate materials with target properties has become a key task in materials science. Machine learning has emerged as an alternative to physics-based simulation, offering a faster and cheaper way to filter materials based on their stability and other target properties, reducing the number of candidates that reach the costly synthesis stage. Recently, Large Language Models (LLMs) have been applied to this role, but these models are parameter-heavy and computationally expensive both during training and at inference time, making them unsuitable for high-throughput tasks. This inefficiency stems from both the large over-parameterization of language models and the difficulty of framing material generation as a sequence learning problem. In this paper, we present PRISMat, a cost-effective, permutation-invariant model, which addresses these limitations. We show that PRISMat, despite taking less time for inference, is able to outperform LLMs in generating crystal slabs conditioned on critical materials' surface properties. In targeted material discovery, we achieve mean absolute errors of 0.188 eV/A

^2

and 2.79 eV for cleavage energy and work function tasks, respectively, reducing the error of the next best model by 4

\times

Claire Schlesinger, Circe Hsu, Peter Schindler +1

Khoury College of Computer Sciences Northeastern University Boston, MA 02115 · College of Engineering Northeastern University Boston, MA 02115

Materials Science Crystal Structure Generation Autoregressive Generation Materials Discovery

Ask PDF

Feb 18, 2026

Clever Materials: When Models Identify Good Materials for the Wrong Reasons

Machine learning can accelerate materials discovery. Models perform impressively on many benchmarks. However, strong benchmark performance does not imply that a model learned chemistry. I test a concrete alternative hypothesis: that property prediction can be driven by bibliographic confounding. Across five tasks spanning MOFs (thermal and solvent stability), perovskite solar cells (efficiency), batteries (capacity), and TADF emitters (emission wavelength), models trained on standard chemical descriptors predict author, journal, and publication year well above chance. When these predicted metadata ("bibliographic fingerprints") are used as the sole input to a second model, performance is sometimes competitive with conventional descriptor-based predictors. These results show that many datasets do not rule out non-chemical explanations of success. Progress requires routine falsification tests (e.g., group/time splits and metadata ablations), datasets designed to resist spurious correlations, and explicit separation of two goals: predictive utility versus evidence of chemical understanding.

Kevin Maik Jablonka

Laboratory of Organic and Macromolecular Chemistry (IOMC), Friedrich Schiller University Jena, Humboldtstrasse 10, 07743 Jena, Germany · Helmholtz Institute for Polymers in Energy Applications Jena (HIPOLE Jena), Lessingstrasse 12-14, 07743 Jena, Germany · Center for Energy and Environmental Chemistry Jena (CEEC Jena), Friedrich Schiller University Jena, Philosophenweg 7a, 07743 Jena, Germany +1

Spurious Correlation Materials Property Prediction Shortcut Learning Materials Discovery Materials Science

Ask PDF

Abstract

Materials Discovery

Explore similar work

Foundation-Model Surrogates Enable Data-Efficient Active Learning for Materials Discovery

PRISMat: Policy-Driven, Permutation-Invariant Autoregressive Material Generation

Clever Materials: When Models Identify Good Materials for the Wrong Reasons