Generative and multimodal AI for materials prediction and design: Progress, challenges, and perspectives

Technical Summary: Generative and Multimodal AI for Materials Prediction and Design

Problem Statement
Artificial intelligence (AI) is increasingly used to navigate vast chemical and structural spaces for materials discovery. However, defining and substantiating "novelty" in materials design remains a critical challenge. Current AI-driven novelty claims often rely on statistical distance in composition or latent space, failing to account for chemical plausibility, physical behavior, and experimental realisability. A significant gap exists between the generation of structurally valid candidates and the delivery of materials that are experimentally feasible and performant under real-world constraints. This disconnect arises because current data and models are heavily concentrated on intrinsic properties (composition and idealized crystal structure) while lacking sufficient evidence from extrinsic, processing-dependent performance data. Furthermore, benchmarking practices often rely on computational proxies (e.g., Density Functional Theory labels) that do not fully reflect experimental deployment conditions, leading to models that may succeed in silico but fail in the laboratory.

Methodology and Framework
The paper proposes a conceptual framework and a review of the current landscape to address these challenges, rather than introducing a specific new algorithm or dataset. The core methodological contributions include:

1. Materials Property Hierarchy: The authors introduce a hierarchy distinguishing between:

   • Intrinsic Properties: Determined by ground-state chemistry and atomic arrangement (e.g., electronic structure), often predictable from idealized structures.
   • Extrinsic Properties: Determined by microstructure, defects, interfaces, and processing history (e.g., fracture toughness, fatigue life). These are path-dependent and require multimodal data to predict.

2. Three-Level Novelty Taxonomy: Building on the property hierarchy, the paper defines three distinct levels of novelty:

   • Structural Novelty: Occupying unexplored regions of compositional or crystallographic space.
   • Physical Novelty: Exhibiting properties or phase behaviors not observed in existing systems under comparable conditions, requiring evidence beyond just structure.
   • Deployment Novelty: Delivering performance unachievable by existing materials under realistic synthesis, manufacturing, and operating constraints. This level demands evidence of synthesisability and retained functionality over time.

3. Multimodal Data Analysis: The paper analyzes the current state of materials data modalities (Chemical Composition, Microstructure, Processing, Testing/Characterisation). It highlights that while composition and idealized structure are well-represented in databases (e.g., Materials Project, OQMD), data regarding processing history, microstructure, and negative results ("dark data") are scarce, heterogeneous, and poorly integrated.

4. Review of AI Architectures: The authors review current approaches in multimodal representation learning (encoding heterogeneous data into shared spaces), knowledge integration (via materials knowledge graphs), and generative AI (including diffusion models, VAEs, and LLMs). They note that while generative models can propose diverse structures, they often lack constraints related to processing feasibility and synthesis routes.

Key Contributions

• Conceptual Clarification: The paper provides a rigorous taxonomy for materials novelty, arguing that structural novelty is insufficient for practical discovery and that higher levels of novelty (physical and deployment) require different combinations of evidence.
• Evidence Gap Identification: It systematically identifies that current AI models are limited by a lack of multimodal data, specifically the scarcity of processing and microstructural data and the systematic absence of negative results in the literature.
• Benchmarking Critique: The paper critiques current benchmarking practices, noting that reliance on computational labels (DFT) and proxy metrics creates a disconnect between model performance and experimental utility. It argues that out-of-distribution evaluation is often insufficiently defined to capture true extrapolation in materials science.
• Strategic Roadmap: The authors outline four specific opportunities to advance the field:
  1. Multimodal Data Construction: Building standardized datasets that link materials to their provenance (synthesis routes, processing conditions, measurement uncertainties) and including negative results.
  2. Process-Aware Modeling: Developing AI models that treat processing information as a primary modality to learn full processing-structure-property relationships, rather than peripheral metadata.
  3. Feasibility-First Generation: Shifting generative AI to embed physical, chemical, and synthesis constraints directly into the generation process, producing "design tuples" (composition + processing route + expected microstructure) rather than isolated structures.
  4. Deployment-Aware Benchmarking: Establishing evaluation protocols that prioritize deployment-relevant metrics (synthesis success rate, constraint satisfaction) and use rigorous out-of-distribution splits that reflect real-world discovery challenges.

Results and Findings
As a perspective paper, the work does not present new experimental results or numerical benchmarks. Instead, its "results" are the synthesis of current trends and the identification of systemic limitations:

• Trend Analysis: Publication data (2020–2025) shows rapid growth in generative and multimodal AI for materials, yet the proportion of papers utilizing all four data modalities (composition, microstructure, processing, testing) remains uneven, with testing and characterisation dominating and processing data remaining under-represented.
• Limitation of Current Models: Current generative models are effective at structural novelty but struggle with physical and deployment novelty because they lack the necessary evidence (processing history, microstructural data) to validate feasibility.
• Benchmarking Deficiencies: Existing benchmarks often reward models for reproducing computational labels or satisfying proxy metrics, which does not guarantee experimental realisability. The gap between computational validation and experimental validity is identified as a fundamental barrier.

Significance and Claims
The paper claims that for AI to successfully design novel materials (rather than rediscover known ones), the community must move beyond structural novelty towards deployment novelty. This transition requires a collective shift in how data is collected, modeled, and evaluated.

• Evidence-Based Novelty: The authors argue that novelty claims must be evidence-based, supported by multimodal data spanning the entire processing-structure-property chain.
• Community Responsibility: Achieving this requires coordinated action across the materials and AI communities to establish minimum metadata standards, enforce the reporting of negative results, and develop infrastructure for aligned multimodal data.
• Future Direction: The paper posits that the future of materials AI lies in "feasibility-first" generative models and "deployment-aware" benchmarks, where the goal is not just to generate valid structures but to propose experimentally realisable materials with defensible scientific and practical novelty.

The paper concludes that while AI has made significant strides in structural prediction, the path to designing experimentally realisable materials depends on closing the evidence gap through better data standards, process-aware modeling, and rigorous, deployment-focused evaluation.