MatExpert: Decomposing Materials Discovery by Mimicking Human Experts

1 INTRODUCTION

• Current material-generation methods often produce static single-step outputs and struggle to integrate property descriptions with structural information for iterative, reasoning-driven design.These limitations constrain generation of materials meeting specific target properties.Source: §1 · p. 2
• MatExpert decomposes material generation into retrieval, transition, and generation stages, using text-structure retrieval and chain-of-thought reasoning to support iterative refinement.The framework is designed to produce more accurate and interpretable material designs aligned with desired specifications.Source: §1 · p. 2
• The work contributes a curated large-scale NOMAD dataset and states that its datasets and source code will be publicly released upon publication.The dataset is presented as a testbed for evaluating MatExpert across diverse material compositions.Source: §1 · p. 2
• MatExpert is evaluated through extensive experiments on the Materials Project and curated NOMAD datasets, supporting assessment across diverse material compositions.The NOMAD dataset contains 2,886,120 materials and is intended as a large-scale testbed for robustness and generalizability.Source: §1 · p. 2
• MatExpert significantly outperforms state-of-the-art baseline methods in material generation experiments.The paper reports this result across both Materials Project and curated NOMAD datasets.Source: §1 · p. 2

2 RELATED WORK

• Existing methods commonly struggle to integrate multimodal data and iteratively refine material designs based on feedback.Most prior work has focused on unconditioned material generation rather than human-expert-like reasoning.Source: §2 · pp. 2–3
• Prior work spans stability-constrained diffusion models, alternative diffusion representations, flow matching, and LLM-based crystal generation from CIF or related structural formats.Examples include CDVAE, fine-tuned LLaMA-2, CrystaLLM, and LM-AC/LM-CH.Source: §2 · pp. 2–3
• Figure 1 reports that MatExpert performs better than baselines across metrics, with especially notable performance on distance metrics.The caption directs readers to Table 1 for detailed results.Source: Figure 1 · p. 2
• The broader literature uses language models for crystal-structure generation, material-property prediction, and structured-data extraction from scientific literature.These applications extend beyond direct material generation into dataset construction and scientific information processing.Source: §2 · p. 3
• MatExpert addresses this shortcoming through multi-stage iterative refinement using retrieval, transition, and generation processes.The framework is presented as an expert-driven alternative to single-stage or unconditioned generation.Source: §2 · p. 3

3 METHODOLOGY

• MatExpert uses retrieval, transition, and generation to design materials through a staged expert-like workflow.The framework combines text-structure retrieval with language-model reasoning and structural generation.Source: §3 · p. 3, Figure 3 · p. 4

3.1 FIRST STAGE: RETRIEVAL

• The retrieval stage identifies the database material that best matches the user’s desired property description.Contrastive learning aligns property and structure embeddings while separating embeddings of other materials.Source: §3.1 · pp. 3–4

3.2 SECOND STAGE: TRANSITION

• The transition stage generates a viable sequence of structural or compositional changes from the retrieved material to the target.GPT-4 supplies ground-truth pathways for fine-tuning a smaller LLM used during practical applications.Source: §3.2 · pp. 4–5

3.3 THIRD STAGE: GENERATION

• The generation stage converts transition information into an ALX description of lattice vectors and atomic coordinates, then into CIF format.The final CIF is produced automatically with pymatgen and contains the predicted material’s structural information.Source: §3.3 · pp. 5–6

4 DATA COLLECTION & PRE-PROCESSING

• The dataset contains 2,886,120 materials after filtering NOMAD entries.The filtering excluded inert elements, non-VASP structures, and structures containing more than 30 atoms.Source: §4 · p. 6
• Each filtered material was converted into CIF format to preserve structural information and support computational-tool integration.CIF is used as the structural representation throughout preprocessing.Source: §4 · p. 6
• The dataset was enriched with RoboCrystallographer descriptions and properties including formation energy, total energy, and elemental composition.M3GNet and PyMatgen were used to extract properties for condition-based design and evaluation.Source: §4 · p. 6
• The curated dataset provides a diverse large-scale basis for training and evaluating MatExpert.The authors describe it as supporting comprehensive assessment across material compositions and properties.Source: §4 · p. 6

5.2 UNCONDITIONAL GENERATION

• MatExpert shows general outperformance over baseline methods for unconditional generation on Materials Project across validity, coverage, property distribution, and stability metrics.Source: Table 1 · p. 7, §5.2 · p. 8

5.3 CONDITIONAL GENERATION

• MatExpert outperforms baseline methods on conditional generation using NOMAD datasets across all reported metrics, including property-distribution distances.Source: Table 2 · p. 7, §5.3 · p. 8

5.2 UNCONDITIONAL GENERATION

• MatExpert outperforms Crystal-LLM across unconditional-generation settings, including matching model size, temperature, and DFT-based stability comparisons.Source: §5.2 · p. 8

5.3 CONDITIONAL GENERATION

• For conditional generation, stability measured by energy above the hull is supplied as an input property rather than separately evaluated as a performance metric.Source: §5.3 · p. 8
• MatExpert consistently achieves higher conditional satisfaction rates than baselines across formula, energy-above-hull, and space-group constraints using 10,000 generated materials per constraint.Source: §5.3 · p. 8, Figure 4 · p. 8

6.1 ANALYSIS OF DIVERSITY AND NOVELTY

• MatExpert maintains high diversity and novelty across structure and composition metrics, while Crystal-LLM’s novelty decreases with larger model size.Source: §6.1 · p. 9, Figure 5 · p. 9

6.2 ABLATION STUDIES

• Removing the transition stage’s Chain-of-Thought reasoning reduces performance, demonstrating the effectiveness of CoT reasoning in MatExpert.Source: §6.2 · p. 9, Figure 6 · p. 9
• Removing retrieval causes a noticeable decline in retrieved-material quality, indicating that high-quality references support subsequent material modifications.Source: §6.2 · p. 9, Figure 6 · p. 9
• T5-based retrieval encoders outperform BERT for property and structure descriptions, producing more accurate embeddings for finding similar materials.Source: §6.2 · pp. 9–10, Table 3 · p. 9

6.3 CASE STUDY

• The conditional-generation case study shows MatExpert taking desired properties as input and generating a material in CIF representation.Source: Figure 7 · p. 10, §6.3 · p. 10

• MatExpert integrates retrieval, transition, and generation stages using LLMs and contrastive learning to design new materials.Source: §7 · p. 11
• Experiments show MatExpert significantly outperforms state-of-the-art methods in material-generation tasks.Source: §7 · p. 11
• The reported results indicate potential for MatExpert as a scalable tool for accelerating material discovery with higher-quality crystal generation.Source: §7 · p. 11

• MatExpert experiments used four NVIDIA A5000 GPUs, a Threadripper PRO 3975WX CPU, and 256 GB RAM for LLaMA-2 and LLaMA-3 8B models.Source: §A.1 · p. 16
• Both machine configurations used Ubuntu 22.04, Python 3.11, PyTorch 2.3.1+cu121, and CUDA 12.4.Source: §A.1 · p. 16
• MatExpert experiments used four NVIDIA L40S GPUs, two Xeon Gold 6526Y CPUs, and 512 GB RAM for LLaMA-2 and LLaMA-3 70B models.Source: §A.1 · p. 16
• The Python library environment requirements were planned for public release alongside MatExpert’s open-source code.Source: §A.1 · p. 16

A.2 IMPLEMENTATION DETAILS

• MatExpert trains T5-based contrastive encoders for retrieval and fine-tunes Llama models for transition and generation.The framework uses separate training stages for retrieval versus transition and generation.Source: §A.2 · p. 16
• Llama-3 8B and 70B use LoRA adaptation, with sequence lengths of 2048 and 1024 tokens, respectively.The 8B model is trained for 70 epochs and the 70B model for 10 epochs; both use gradient accumulation over 8 steps.Source: §A.2 · p. 17

A.3 EFFICIENCY

• Fine-tuning takes around 30 hours for Llama-3 8B and 120 hours for Llama-3 70B.The authors use LLaMA-Factory with LoRA and FlashAttention for fine-tuning.Source: §A.3 · p. 17
• MatExpert with Llama-3 8B requires ∼40 minutes per 10,000 samples, while with Llama-3 70B it requires ∼5 hours per 10,000 samples.Source: §A.3 · p. 17

B DATASETS DETAILS

• Materials Project and NOMAD provide the datasets for unconditional and conditional generation, respectively.Materials Project is also called MP-20 and contains 45,231 materials, while NOMAD contains 2,886,120 materials.Source: B DATASETS DETAILS · p. 17
• NOMAD data exclude inert elements and retain structures simulated using VASP.The collection uses NOMAD’s API and requests structural properties and reduced chemical formulas.Source: B DATASETS DETAILS · pp. 17–18
• The workflow converts structure data from JSON into CIF files using lattice parameters or vectors, species, and atomic positions.The conversion constructs a pymatgen Structure before producing the CIF representation.Source: B DATASETS DETAILS · p. 18
• Two random sets of 10,000 samples serve as validation and testing data, with all remaining samples used for training.The same split procedure is stated for the collected materials.Source: B DATASETS DETAILS · p. 18

C.1 VALIDITY

• Validity requires interatomic distances above 0.5 Å and neutral overall charge.These criteria assess structural integrity and chemical feasibility.Source: §C.1 · p. 18

C.2 COVERAGE

• Coverage uses COV-R and COV-P to measure resemblance to ground-truth materials, computed over 10,000 generated samples.COV-R measures recall of ground-truth materials, while COV-P measures precision of generated material diversity.Source: §C.2 · p. 18

C.3 DISTRIBUTION

• Wasserstein distance compares generated and test-set distributions for density and the number of unique elements over 1,000 valid materials.The sampled materials must first pass the validity test.Source: §C.3 · p. 19

C.4 DIVERSITY

• Diversity uses pairwise structural and compositional distances among M3GNet-metastable samples, normalized to the test set.The metric quantifies how distinct generated samples are from one another.Source: §C.4 · p. 19

C.5 NOVELTY

• Novelty compares each generated sample with its nearest training-set neighbor using structural and compositional cutoffs of 0.1 and 2.A crystal is novel when its nearest neighbor exceeds either predefined distance threshold.Source: §C.5 · p. 19

C.6 STABILITY

• Stability evaluation combines M3GNet relaxation with DFT relaxation under Materials Project-compatible settings.M3GNet identifies metastable candidates, while DFT further evaluates those candidates using hull-energy thresholds below 0.1 and 0.0.Source: §C.6 · p. 19

D CASES OF UNCONDITIONAL GENERATION

• Figures 8 and 9 present cases of unconditional and conditional generation, respectively.The supplied passages identify the figure scopes but do not report a specific comparison or outcome.Source: Figure 8 · p. 20, Figure 9 · p. 21