aLLoyM: LLM for Alloy Phase Diagrams

• aLLoyM is a domain-specific large language model designed for alloy phase diagram prediction using CALPHAD-derived Q&A data.
• It employs a fine-tuned Mistral-based system with LoRA adapters to provide phase-equilibrium guidance in both prediction and generative settings.
• Its standardized phase nomenclature supports closed-loop experimental design by identifying composition regions with high phase complexity.

aLLoyM is a LLM specialized for alloy phase diagrams. It was introduced as a fine-tuned Mistral-based system trained on alloy compositions, temperatures, and corresponding phase information derived from the Computational Phase Diagram Database (CPDDB), with the explicit aim of treating phase-diagram reasoning as a question–answer problem over composition and temperature rather than as a narrowly defined regression task (Oikawa et al., 30 Jul 2025). In subsequent work, aLLoyM was used as a domain-specific prior inside a closed-loop experimental framework for phase diagram construction, where its principal value was not exact phase-name correctness but guidance toward composition regions where complicated multiphase equilibria were likely (Tamura et al., 22 Apr 2026).

1. Definition and domain role

aLLoyM is a domain-specific LLM for alloy phase diagram prediction. In the 2025 formulation, it is described as a LLM specialized for alloy phase diagrams and trained on CALPHAD-derived question-and-answer pairs; in the 2026 experimental-planning study, it is described as a domain-specific LLM trained specifically to predict equilibrium phases in alloy systems, in contrast to a general-purpose LLM such as Claude (Oikawa et al., 30 Jul 2025).

Its operational premise is that phase diagrams can be represented as structured Q&A mappings. The model is intended to learn associations of the form composition plus temperature to phase name(s), composition plus temperature to full phase information, and element set plus phase domain to plausible composition and temperature. This formulation makes aLLoyM usable both as a predictor and as a generative tool. A central implication is that the model occupies an intermediate position between thermodynamic databases and open-ended LLM reasoning: it is not a replacement for CALPHAD assessment, but a fine-tuned assistant that can answer phase-diagram queries and propose hypotheses in unexplored chemical spaces (Oikawa et al., 30 Jul 2025).

The later experimental study further narrows this role. There, aLLoyM is not treated as a full autonomous planner. Instead, it serves as a specialized prior that supplies approximate prior information about likely phase equilibria at a given composition and temperature, while a general-purpose LLM acts as the experimental planner in the closed loop (Tamura et al., 22 Apr 2026).

2. Training corpus and task construction

The training corpus for aLLoyM was built from the open-source CPDDB repository from NIMS. The authors extracted thermodynamic database files for 389 binary phase diagrams and 38 ternary phase diagrams, then used Pandat to perform CALPHAD calculations on systematic grids. For binary systems, compositions were sampled from 0% to 100% in 2% increments and temperatures from 200 K to 5000 K in 50 K steps. For ternary systems, temperature was fixed at 800 K. This produced 837,475 data points linking elemental composition, temperature, and phase information (Oikawa et al., 30 Jul 2025).

These data were then converted into question-and-answer examples in a unified framework with three tasks. The first task, full phase information, uses composition and temperature as input and returns full phase description, including phase names, fractions, and phase compositions. The second task, phase name, uses the same input but returns only the phase names. The third task, experimental condition, inverts the direction: it uses element set plus phase domain as input and returns plausible composition and temperature. Training examples were formatted with a structured prompt template containing Instruction, Input, and Output (Oikawa et al., 30 Jul 2025).

This task formulation is significant because it extends phase-diagram modeling beyond direct forward prediction. The model is trained not only to recognize equilibrium phase outputs, but also to infer experimental conditions from desired phase domains. This suggests a bidirectional use of phase-diagram information, albeit within the limits of CALPHAD-derived supervision rather than experimentally complete thermodynamic truth (Oikawa et al., 30 Jul 2025).

3. Base model, fine-tuning, and output conventions

The base model for the 2025 system was Mistral-Nemo-Instruct-2407-bnb-4bit from Hugging Face. aLLoyM was fine-tuned using LoRA (Low-Rank Adaptation) with rank = 16, alpha = 16, and target modules in the attention and feed-forward projections. The reported training hyperparameters were 15,000 steps, learning rate = $2 \times 10^{-4}$, batch size = 16 per device, 4 gradient accumulation steps, optimizer = AdamW, and precision = bfloat16. The 2026 study characterizes the deployed model as a Mistral-Nemo-Instruct (12B parameter) model fine-tuned with LoRA adapters on the Computational Phase Diagram Database (CPDDB), which is consistent with the original description of a Mistral-based LoRA-adapted system (Oikawa et al., 30 Jul 2025).

A noteworthy feature of aLLoyM in experimental deployment is its output vocabulary. Its outputs are not system-specific phase names, but a standardized phase nomenclature based on structural prototypes and generic labels such as B2, C36, FCC_A1, SOLID, and LIQUID. This means that the model does not attempt to name the exact experimental phase that will appear. Instead, it provides approximate prior information about what kinds of phase equilibria are likely at a given composition and temperature. In that sense, aLLoyM is optimized for phase-domain guidance rather than for direct reproduction of experimental naming conventions (Tamura et al., 22 Apr 2026).

This output design matters methodologically. Because the labels are standardized and CALPHAD-derived, the model’s predictions are best interpreted as structurally typed phase-diagram priors. A plausible implication is that this nomenclature stabilizes training across many systems while also constraining how literally predictions should be mapped onto experimentally resolved phase identities (Tamura et al., 22 Apr 2026).

4. Evaluation protocols and quantitative scoring

aLLoyM was evaluated in two formats: multiple-choice Q&A and short-answer Q&A. In the multiple-choice setting, each question had four options, comprising 1 correct answer and 3 distractors randomly chosen from answers related to the same systems. The dataset was split 8:2 into training and test sets under two regimes: an interpolation split, where points were randomly distributed across all systems, and an extrapolation split, where entire systems were held out from training. These evaluations covered the same three task types: full phase information, phase name, and experimental condition (Oikawa et al., 30 Jul 2025).

The baseline was the untuned Mistral-Nemo-Instruct-2407-bnb-4bit model. The reported result is qualitative but unambiguous: the baseline performed near random guessing in both interpolation and extrapolation settings, whereas fine-tuning substantially improved performance across all tasks. Performance was better on interpolation than extrapolation, better on binary than ternary systems, and showed only minor differences among the three task types in the multiple-choice setting. For short-answer generation, the model again performed better on interpolation than extrapolation, with stronger results for phase-name prediction than for full phase information and robust phase-name prediction even on extrapolated systems (Oikawa et al., 30 Jul 2025).

The short-answer evaluation used task-specific metrics. For full phase information, the score was exact-match only:

$$\text{Score} = \begin{cases} 100\% & \text{if the generated answer exactly matches the ground truth} \ 0\% & \text{otherwise} \end{cases}$$

For phase name, predicted and true phase-domain lists $A$ and $B$ were scored by Jaccard similarity:

$$\text{Score} = \begin{cases} |A \cap B| / |A \cup B| \times 100\% & \text{if } |A \cup B| > 0 \ 0\% & \text{otherwise} \end{cases}$$

with

$$A = [\text{phase name }1, \ldots, \text{phase name }k], \quad B = [\text{phase name }1, \ldots, \text{phase name }l].$$

For experimental condition, the paper defined

$$A_{\rm c} = \frac{1}{N} \sum_{n = 1}^{N} \left( 1 - \frac{|x_n^* - x_n|}{100} \right),$$

$A_{\rm t} = 1 - \frac{|T^* - T|}{\Delta T},$

and

$$\text{Score} = \left( \frac{1}{2} A_{\rm c} + \frac{1}{2} A_{\rm t} \right) \times 100\%.$$

If multiple ground truths existed, the score was computed against each and the maximum retained. These scoring rules are important because they formalize partial correctness for free-form phase and condition generation rather than requiring only categorical exact matches (Oikawa et al., 30 Jul 2025).

5. Generative behavior, extrapolation, and limitations

A central claim of the 2025 paper is that the short-answer model can generate novel phase diagrams from its components alone. The authors report examples for Th–Ac, U–Nh, W–Ta–Os, and Nh–U–Ac. These examples include partially known and completely hypothetical systems. For Th–Ac, actinium was not in the training data, but the model predicted a melting point roughly consistent with experiment while mispredicting the stable crystal structure. For U–Nh, neither element was in training data, yet the model produced a plausible phase diagram and a moderate estimate of uranium’s melting point, again with structural errors. For W–Ta–Os, no ternary diagram was known experimentally, but the model generated an 800 K ternary isothermal section and predicted phases with “WOLF” nomenclature absent from training data. For Nh–U–Ac, it produced a hypothetical ternary diagram for a system that cannot be experimentally realized (Oikawa et al., 30 Jul 2025).

These results are presented with explicit caveats. The paper states that many outputs remain beyond experimental validation and that some predictions are chemically incorrect, so the short-answer model is best seen as a generative hypothesis engine rather than a validated predictor. The model tends to perform better near pure elements and worse in intermediate composition regions, especially where phase behavior is complex. It also often misidentifies crystal structures, and generated “novel” diagrams tend to reuse phase names seen during training, which indicates limited novelty in phase vocabulary. Ternary performance is weaker than binary performance because there is less ternary training data, and the authors identify more ternary and higher-order training data, better prompt engineering, thermodynamics-aware prompting, and better control over generated phase names as future directions (Oikawa et al., 30 Jul 2025).

The paper also emphasizes reproducibility and reuse. It states that the short-answer fine-tuned aLLoyM and the complete benchmarking Q&A dataset are publicly released on Hugging Face, and that code is available on GitHub. This release is positioned as enabling reproducibility, benchmarking, reuse of the dataset for further model development, and community-driven improvements for phase-diagram prediction (Oikawa et al., 30 Jul 2025).

6. Experimental use in LLM-guided phase diagram construction

aLLoyM’s most concrete downstream use, within the provided literature, is in the 2026 framework for LLM-guided phase diagram construction through high-throughput experimentation. The overall workflow was a closed loop: the AI planner selected compositions, samples were synthesized by high-throughput methods, X-ray diffraction identified the phases, and the new results were fed back into the planner for the next cycle. This loop was repeated for six cycles, with eight compositions per cycle for a total of 48 measurements per strategy. The experimental system was the Co-Al-Ge ternary at 900 °C (1173.15 K) with 231 candidate compositions on a 5 at.% grid (Tamura et al., 22 Apr 2026).

Two strategies were compared:

Strategy	Initial planning	Subsequent planning
Strategy A	aLLoyM-guided initial cycle	general-purpose LLM from cycle 2 onward
Strategy B	general-purpose LLM only	general-purpose LLM only

In both strategies, the general-purpose LLM was run as an agent via Claude Code, and each cycle used 10 independent selection runs, with the final 8 points chosen by majority vote. The difference between the strategies was therefore mainly in the starting bias. Strategy A used aLLoyM predictions as reference information in the first cycle, which pushed selections toward the interior of the ternary triangle; Strategy B followed a more textbook-like exploration that sampled the corners and binary edges before moving inward (Tamura et al., 22 Apr 2026).

This distinction had experimentally measurable consequences. aLLoyM-guided points tended to target compositionally complex interior regions, and this enabled the earliest discovery of all three novel ternary-only phases. Specifically, B20. Co(Al/Ge) was found in cycle 1 by Strategy A but only in cycle 3 by Strategy B; X was found in cycle 2 by Strategy A but only in cycle 4 by Strategy B; Co2(Al/Ge)3 was found by both strategies in cycle 1. Most importantly, Strategy A discovered all three ternary-only novel phases within 16 measurements (cycle 2), whereas Strategy B needed 32 measurements (cycle 4). The paper gives examples of aLLoyM-predicted labels that nonetheless led to discovery: a point predicted as C36 led to Co2(Al/Ge)3; a point predicted as C30 + FCC_A1 led to B20. Co(Al/Ge); a point predicted as CQG_B2 again led to Co2(Al/Ge)3; and a point predicted only as SOLID was selected because of its ambiguity and yielded B20. Co(Al/Ge) (Tamura et al., 22 Apr 2026).

The complementary strength of the general-purpose LLM also remained clear. Strategy B was faster at broad phase coverage, discovering 9 phases by cycle 2 and all 11 phases by cycle 4, whereas Strategy A reached 11 phases only by cycle 6. The paper therefore does not frame the result as a narrow “aLLoyM vs Claude” comparison. Instead, it argues that aLLoyM = domain-specialized phase predictor and general-purpose LLM = flexible experimental planner, with the combination being particularly effective when the objective is novelty discovery rather than exhaustive mapping. The “phase diagram” constructed in that study is explicitly not a thermodynamic equilibrium phase diagram in the strict sense; it is the experimentally obtained primary phase assemblage map after heating powders at 900 °C and cooling. Within that experimental framing, aLLoyM’s distinctive contribution was to identify regions of high phase complexity rather than to supply exact phase identities (Tamura et al., 22 Apr 2026).