MatVerse: A Foundation Model for Materials Science
Authors: Xuanyu Zhu, Qian Tan, Bo Chen, Jiaqing Xie, Yang Han, Mei Wu, Wanhao Liu, Xuanze Lin, Yuchen Zhou, Shijie Wang, Biqing Qi, Shixiang Tang, Wenwei Zhang, Tianfan Fu, Kai Yu, Lu Chen, Yuqiang Li, Xin Chen
Organizations: Suzhou Laboratory · Shanghai Artificial Intelligence Laboratory · Shanghai Jiao Tong University · University of Science and Technology of China · Fudan University · Nanjing University
Materials science involves highly heterogeneous scientific information, including textual knowledge, molecular and crystal structures, experimental characterization data, microscopy images, and structure-property relationships. Developing a general-purpose foundation model capable of understanding and reasoning over these diverse materials representations remains challenging. In this technical report, we introduce MatVerse, a materials foundation model built upon Intern-S2-Preview and further specialized through large-scale materials-domain post-training. We construct a large-scale multimodal training corpus covering materials knowledge, scientific reasoning, quantitative problem solving, structural understanding, and multimodal scientific interpretation. To improve domain adaptation, we develop a unified data construction pipeline combining teacher-guided reasoning distillation, scientific problem synthesis, consistency-based filtering, and retrieval-based multimodal data construction. We evaluate MatVerse on three representative materials-science benchmarks: our human-calibrated OmniMatBench, MatSciBench, and MaCBench. Results demonstrate consistent improvements after materials-specific post-training over the base model on all three benchmarks, alongside improved inference efficiency on calculation-intensive problems. These results show that targeted domain-specific post-training can effectively enhance scientific reasoning, materials knowledge understanding, and multimodal scientific capabilities of foundation models. MatVerse provides a unified language-and-vision interface for materials science and serves as a foundation for future AI-assisted materials research workflows.
Figures & tables
Figure 1: Overall comparison of MatVerse against representative foundation models on MatSciBench, MaCBench, and the OmniMatBench CAL track. On every benchmark, MatVerse substantially improves over its base model Intern-S2-Preview (35B) and achieves performance competitive with much larger frontier proprietary models, demonstrating that materials-specific post-training yields strong gains in scientific reasoning, multimodal scientific understanding, and quantitative calculation.
Figure 2: Overview of MatVerse . (a) Materials-domain specialization: reasoning-oriented (30,639 examples) and multimodal (119,153 examples) supervision streams are merged into 149,792 SFT examples. The data and SFT recipe are validated on a Qwen3-VL-4B-Thinking pilot model before scaling, and the final model is obtained by materials-domain SFT of Intern-S2-Preview-35B-A3B, which inherits its multimodal architecture from the Qwen3.5 35B-A3B language backbone. (b) Unified text–image interface: MatVerse accepts natural language and symbolic representations together with optional scientific imagery, and supports diverse materials-science tasks within a shared backbone.
Figure 3: Materials data construction for MatVerse. The reasoning-oriented stream synthesizes new problems from seed examples and applies teacher-guided solution generation and quality control according to answer availability. The multimodal stream uses text-to-image and image-to-image retrieval to select scientific images for generating image-grounded questions, answers, and short explanations. The two streams contribute 30,639 and 119,153 examples, respectively, yielding 149,792 post-training examples in total.
Data category
# Samples
Ratio
Broad materials reasoning and synthesis
27,101
18.1%
Quantitative materials reasoning
3,538
2.4%
Multimodal multiple-choice data
111,598
74.5%
Multimodal open-ended data
7,555
5.0%
Total
149,792
100%
Table 1: Composition of the materials post-training corpus.
Benchmark
Metric
Base / Ref.
MatVerse
Δ
MatSciBench
rule accuracy
58.45%
63.59%
+5.14
MaCBench
overall accuracy
65.05%
67.22%
+2.17
OmniMatBench (CAL)
ESA
21.98%
24.45%
+2.47
OmniMatBench (QA)
Macro-F 1
28.67%
27.59%
-1.08
Table 2: Overall primary-metric results on the three benchmarks.
Figure 4: Fine-grained results of MatVerse (Intern-S2-Preview in pale bars, MatVerse in blue, hatched in a–b). (a) MatSciBench breakdown by problem type and modality. (b) MaCBench breakdown by capability tag. (c) OmniMatBench CAL (exact slot accuracy) and (d) QA (Macro-F 1 ) compared with representative foundation models. Models with diagonally hatched bars are proprietary systems accessed via API; all other models, including MatVerse , are open-weight and available on Hugging Face.
Figure 5: MaCBench task-family results (overall accuracy, %; Intern-S2-Preview vs. MatVerse ). Regressed task families are highlighted in red.
ID
Subfield
CAL #Q
ESA
Compl.
QA #Q
Macro-F 1
01
Materials Science and Engineering
78
36.32%
100%
85
12.10%
02
Soft Matter Science and Engineering
64
17.26%
98.44%
70
33.69%
03
Materials Physics
91
30.00%
97.80%
75
28.99%
04
Materials Chemistry
96
41.14%
97.92%
80
48.29%
05
Materials Design Science and Engineering
87
24.17%
98.85%
78
36.77%
06
Metallurgical Engineering
86
1.30%
98.84%
96
12.89%
Table 3: MatVerse on OmniMatBench by subfield. CAL reports exact slot accuracy and the completion rate; QA reports Macro-F 1 .
Figure 6: Data recipe, domain gap, and scale analysis. (a) OmniMatBench CAL exact slot accuracy of the Qwen3-VL-4B-Thinking pilot as the data recipe progresses from the base model through rejection-sampling (RS) distillation data v1/v2 to the full materials mixture. (b) exact slot accuracy of chemistry-specific models, the 35B base model, and MatVerse on OmniMatBench CAL. (c) MatSciBench rule accuracy across model scales, including the 397B Intern-S2-Official baseline.
Figure 7: Qualitative question–answering examples of MatVerse sampled from the MaCBench evaluation (responses condensed; the model answers all six questions correctly). (a) Hand-drawn molecule to SMILES. (b) Laboratory equipment recognition. (c) Crystal-system identification from a structure rendering. (d) Data extraction from a US-patent plot. (e) Materials property table understanding. (f) XRD relative-intensity analysis. Full prompts, thinking trajectories, and complete responses are provided in Appendix D.8 ; condensed long-chain MatSciBench reasoning cases are provided in Appendix D.9 .
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Statistic
Value
Synthetic questions before filtering
5,938
Samples passing majority filtering
2,240
Samples rejected by consistency filtering
3,512
Intermediate retention rate
37.7%
Samples with non-empty reasoning
100%
Appendix
Table 4: Statistics of multi-rollout consistency filtering.
Statistic
Value
Unique images retrieved by text-to-image search
304,504
Unique images retrieved by image-to-image search
263,443
Images retrieved by both pathways
86,749
Same image retrieved for the same query by both pathways
18,792
Retrieval depth per pathway
200
Appendix
Table 5: Full dual-path retrieval pool. Counts are unique images, except the last overlap row, which counts query–image pairs.
Figure 8: Image-level Venn diagram of dual-path retrieval. Text-to-image and image-to-image return 304,504 and 263,443 unique images, respectively; their intersection contains 86,749 images ( 217,755 text-only and 176,694 image-only). The bars below report the within-pathway shares of the intersection. Same-query both-pathway matches ( 18,792 ) are listed in Table 5 and are not drawn here.
Statistic
Value
Both-pathway records linked to a MaCBench query
52,922
Unique images in that export
52,902
Records with score ≥0.60
52,246
Records removed by the score cutoff
676
Images with a successful generation
51,757
Question–answer pairs produced
119,167
Appendix
Table 6: Statistics of the both-pathway subset used for multimodal question–answer synthesis.
Figure 9: Similarity-score distribution of the 52,922 both-pathway records linked to MaCBench queries. The pale bars indicate the 676 records (1.28%) below the 0.60 synthesis threshold, which are excluded before question–answer generation. Image-only matches with scores above 0.85 are not included in this subset.
Hyperparameter
Value
Post-training objective
Supervised fine-tuning (SFT)
Trainable components
Language model (full parameter)
Frozen components
Visual encoder, multimodal alignment modules
Optimizer
Distributed AdamW-style optimizer
Peak learning rate
1×10−5
Minimum learning rate
1×10−6
Appendix
Table 7: Hyperparameter settings for the 4B pilot training of MatVerse .
As multimodal language models play an increasingly important role in scientific research, materials science offers a critical testbed due to its interdisciplinary, multimodal, and application-driven nature. However, existing materials benchmarks mainly focus on property prediction, knowledge QA, or characterization understanding, leaving the broader reasoning process from materials knowledge to application underexplored. To fill this gap, we present OmniMatBench, a human-calibrated multimodal reasoning benchmark for materials science. OmniMatBench contains 3,171 expert-curated QA and calculation problems across 19 materials-science subfields, spanning fundamental materials knowledge, structural and engineering materials, materials processing and manufacturing, and functional and applied materials. We evaluate 13 open-source and closed-source MLLMs and find that the best model achieves only a 0.372 overall score, revealing a substantial gap in current materials-science reasoning. Further analysis shows strong variation across subfields, fixed reasoning heuristics, uneven materials knowledge, and limited high-level knowledge application under formula-, retrieval-, and code-assisted settings. OmniMatBench provides crucial insights into the capabilities and limitations of current MLLMs and establishes a foundation for reliable AI assistants in materials-science research.
Wanhao Liu, Jiaqing Xie, Qian Tan +10
University of Science and Technology of China · Shanghai Artificial Intelligence Laboratory · Fudan University +4
Large language models and the multi-agent systems built around them could accelerate materials discovery by connecting scientific knowledge with experimental decisions. Existing benchmarks establish capabilities in domain knowledge, multimodal interpretation and scientific tool use. As language models increasingly support materials research workflows, their ability to distinguish closely related material states and reason accurately with quantitative evidence remains insufficiently evaluated. These abilities are essential for comparing material responses, assessing mechanisms and maintaining reliable judgements as experimental evidence accumulates. We introduce MatScope to evaluate three connected capabilities. M1 tests unaided, condition-specific reasoning through 200 tasks spanning ten research functions, using aluminium and high-entropy alloys as representative material systems. M2 tests agent-assisted transfer of related knowledge to unresolved materials questions. MatEvoBench-20 tests initial forecasts across scientific dependency chains and revisions of still-hidden responses as observations accumulate. Across eleven model deployments, including GPT-6 Astra and Claude Fable 5.1, given-input calculations and many physical explanations are handled well. Fine-grained response estimates, comparisons between similar states and quantitatively supported choices remain substantially weaker. Related literature improves selected estimates, while condition and baseline mismatches limit transfer. Progressive observations calibrate some downstream responses and destabilize earlier accurate predictions. MatScope establishes measurable targets for developing reliable scientific reasoning in future self-evolving materials laboratories and materials-aware world models.
Xuyu Dong, Yixuan Li, Jianguo Huang +5
Shanghai Key Lab of Advanced High-temperature Materials and Precision Forming and State Key Lab of Metal Matrix Composites, School of Materials Science and Engineering, Shanghai Jiao Tong University, Shanghai 200240, China. · School of Materials Science and Engineering, Nanyang Technological University, Singapore, Singapore. · College of Computing and Data Science, Nanyang Technological University, Singapore, Singapore.
To accelerate materials discovery, language models must be able to address problems at the research frontier. Yet textbook- and examination-derived benchmarks primarily assess established knowledge, leaving performance on frontier research problems largely untested. We introduce FrontierMat, a benchmark of 330 problems drawn from frontier academic and industrial research in fields including aerospace, semiconductors, embodied intelligence and nuclear fusion. Every problem is validated by leading experimentalists and industry experts in its field. Each problem presents a real study and asks the model to infer a withheld outcome or draw a conclusion from selected experimental evidence. Problems are annotated by functional domain, material system and the scale of the governing mechanism; 130 require cross-scale reasoning. The strongest general-purpose model scores only 53.7%, and on matched problems, cross-scale problems score 8.7 points below single-scale ones. Half of the problems are solved by at most one model, and models fail chiefly by endorsing what is plausible rather than what was observed. A domain-adaptation study with meta-weighted continued pretraining and scale-modular self-distillation raises an open-weight model from 36.7% to 55.4%, matching the strongest proprietary model, with improvements on public materials benchmarks and no decline in aggregate performance on general benchmarks. FrontierMat provides a testbed for measuring progress toward language models that support frontier materials R&D.
Feng Chen, Jianguo Huang, Yushan Xiao +20
Nanyang Technological University, Singapore · Peking University · Shanghai Jiao Tong University +11