Domain-Informed AI Multimodal Foundation Model for Materials Science
Abstract
Accelerating materials discovery can unlock progress on clean energy, sustainable construction, medicines, and electronics, yet the machine-learning pipelines meant to drive that acceleration remain fragmented, unimodal, and narrowly scoped. Large multimodal foundation models, pretrained broadly and reused across tasks, offer an alternative paradigm. Materials-specific foundation models already exist and have delivered genuine advances, most notably universal interatomic potentials and domain language models, but those surveyed here remain either unimodal or physics-agnostic, and building a model that reasons jointly across the full range of materials data requires more than scaling them up. This review examines what such a model would demand in order to deliver reliable, physically consistent property prediction and candidate generation that transfer across chemistries, length scales, and data regimes, and how close the field is to meeting those demands. We survey the data infrastructure and unimodal precursors already in place, analyze the recent landscape of foundation models for materials, and argue that a credible model must be domain-informed; because most materials knowledge is formalized as physical laws, we concentrate on its physics-informed subset. Organized around four pillars, we review multimodal data integration, model architecture and pretraining, parameter-efficient specialization to low-data industrial problems, and validation, deployment, and continuous improvement. A recurring finding is that physical grounding and multimodality have so far advanced on separate lines. Beyond reviewing this landscape, we use the state-of-theart analysis to propose guidelines for further developments, arguing that a physically grounded multimodal backbone would let discovery in data-rich sectors transfer to data-poor ones.
Related
- ChemEidos: Synthesis-Aware Multimodal Foundation Models for Domain-Specific Generative Design
- A Survey of AI for Materials Science: Foundation Models, LLM Agents, Datasets, and Tools
- Realistic Simulations of Energy Materials Using Foundation Models and Electrode-Potential Learning
- Chemistry-Informed Transfer and Interaction-Resolved Learning Towards Data-Efficient Materials Exploration
- Correlative SEM–AFM Characterization of 2D Materials: From Sequential Measurements to In-Situ Multimodal Platforms
BibTeX
@misc{rebello2026domaininformed,
title = {{Domain-Informed AI Multimodal Foundation Model for Materials Science}},
author = {Carine de Menezes Rebello and Erick Giovani Sperandio Nascimento and Idelfonso B. R. Nogueira},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15008789/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008789/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Domain-Informed AI Multimodal Foundation Model for Materials Science
AU - Carine de Menezes Rebello
AU - Erick Giovani Sperandio Nascimento
AU - Idelfonso B. R. Nogueira
PY - 2026
DA - 2026/09/14
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008789/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008789/v1
AB - Accelerating materials discovery can unlock progress on clean energy, sustainable construction, medicines, and electronics, yet the machine-learning pipelines meant to drive that acceleration remain fragmented, unimodal, and narrowly scoped. Large multimodal foundation models, pretrained broadly and reused across tasks, offer an alternative paradigm. Materials-specific foundation models already exist and have delivered genuine advances, most notably universal interatomic potentials and domain language models, but those surveyed here remain either unimodal or physics-agnostic, and building a model that reasons jointly across the full range of materials data requires more than scaling them up. This review examines what such a model would demand in order to deliver reliable, physically consistent property prediction and candidate generation that transfer across chemistries, length scales, and data regimes, and how close the field is to meeting those demands. We survey the data infrastructure and unimodal precursors already in place, analyze the recent landscape of foundation models for materials, and argue that a credible model must be domain-informed; because most materials knowledge is formalized as physical laws, we concentrate on its physics-informed subset. Organized around four pillars, we review multimodal data integration, model architecture and pretraining, parameter-efficient specialization to low-data industrial problems, and validation, deployment, and continuous improvement. A recurring finding is that physical grounding and multimodality have so far advanced on separate lines. Beyond reviewing this landscape, we use the state-of-theart analysis to propose guidelines for further developments, arguing that a physically grounded multimodal backbone would let discovery in data-rich sectors transfer to data-poor ones.
ER -