Physics adapted generative AI for metal insulator transition materials under label scarcity

Technical Summary: Physics-Adapted Generative AI for Metal-Insulator Transition Materials Under Label Scarcity

Problem Statement
The discovery of Metal-Insulator Transition (MIT) materials is hindered by a fundamental mismatch between data availability and the complexity of the target phenomenon. While large first-principles databases contain abundant information on stable inorganic structures, verified MIT labels are extremely scarce (e.g., only 68 confirmed MIT compounds in a curated database of 343 entries). Unlike standard materials discovery tasks that optimize static properties (e.g., formation energy or band gap), MIT discovery requires identifying a switchable phase landscape. A valid MIT candidate must possess at least two electronically distinct states (metallic and insulating) connected by an experimentally accessible control parameter (temperature, pressure, strain, field, etc.) within a useful operating window.

Current generative AI models in materials science are primarily trained to generate stable crystals or predict single-point properties. They struggle with MIT discovery because:

1. Label Scarcity: There is insufficient data to train models directly on the MIT class.
2. Mechanism Complexity: MITs arise from diverse, often intertwined mechanisms (Mott-Hubbard, Peierls distortion, charge ordering, excitonic instability, etc.) rather than a single universal rule.
3. Context Dependence: A material's MIT behavior is highly sensitive to stoichiometry, dimensionality, and external conditions, making binary "MIT/non-MIT" labels ambiguous and device-relevance difficult to assess.

Methodology: A Modular Physics-Adapted Framework
The authors propose a "physics-adapted" generative framework designed to operate under label scarcity by shifting the objective from generating a "discovered material" to generating a "mechanism-informed phase-transition hypothesis." The approach is modular:

1. Pretrained Crystal Generator: A foundation model (e.g., a crystal diffusion model) pretrained on large-scale inorganic datasets provides the broad structural prior, ensuring generated candidates are chemically valid and crystallographically realistic.
2. Domain Adaptation: A lightweight adapter biases the generation toward specific chemical families known to host correlated physics, specifically transition-metal oxides and chalcogenides.
3. Mechanism-Prior Adapters: Instead of training a classifier on the small set of known MITs, the framework uses "mechanism-prior adapters." These are controllable biases steered by scalable physics proxies (e.g., bandwidth-to-Coulomb repulsion ratios, metal-metal distances, orbital degeneracy, or phonon softening tendencies) rather than rigid mechanism labels. These adapters guide the generator toward specific physical routes (e.g., dimerization, charge disproportionation) without requiring retraining of the base model.
4. Parameter-Efficient Adaptation: Techniques like Low-Rank Adaptation (LoRA) are suggested to implement these biases efficiently, allowing the model to explore different physical hypotheses without full retraining.

Verification Ladder
The paper emphasizes that generative output must be treated as a hypothesis subject to a staged "verification ladder" before being considered a credible candidate. The authors outline six progressive stages:

1. Structural/Chemical Validity: Sensible interatomic distances and oxidation states.
2. Thermodynamic Accessibility: Quantitative assessment of energy above the convex hull ($E_{hull}$) and metastability windows.
3. Phase Competition: Verification that the candidate supports a plausible competing state (e.g., a distorted phase, magnetic variant, or charge-ordered arrangement) with an accessible energy separation (ideally $\sim$25–50 meV per formula unit).
4. Electronic Contrast: Demonstration of a meaningful change in electronic character (e.g., gap opening, suppression of density of states at the Fermi level, $N(E_F)$).
5. Pathway Plausibility: Identification of the physical mechanism (e.g., soft phonon modes, symmetry breaking) connecting the phases.
6. Experimental/Device Accessibility: Assessment of the transition driver, operating window, hysteresis, and integration feasibility.

Key Contributions

• Reframing the Objective: The paper argues that MIT discovery should be framed as the inverse design of switchable phase landscapes rather than the generation of static stable crystals.
• Modular Architecture: It proposes a specific architecture combining a broad pretrained generator with domain and mechanism-prior adapters to navigate the scarcity of MIT labels.
• Mechanism-Prior Steering: It introduces the concept of using scalable physics proxies to steer generation toward specific instability types (e.g., Mott, Peierls, CDW) rather than relying on supervised learning of mechanism classes.
• Verification-Centric Roadmap: It establishes a rigorous, multi-stage verification protocol that shifts the focus from "novelty" to "hypothesis credibility," defining quantitative gates for stability, phase separation, and electronic contrast.

Results and Claims
As a "Perspective" paper, this work does not present new experimental data or a specific trained model with benchmark results. Instead, its "results" are conceptual and strategic:

• It demonstrates that existing large-scale datasets and generative models are insufficient for MIT discovery without specific physics-adapted modifications.
• It illustrates, through examples like $VO_2$, $1T-TaS_2$, and $Ta_2NiSe_5$, how a mechanism-informed approach can better handle the ambiguity and complexity of real-world MIT systems compared to binary classification.
• It claims that the proposed framework can improve the "front end" of discovery by enriching the pool of generated candidates that survive MIT-specific verification, rather than claiming to directly discover new materials.

Significance
The paper's significance lies in its call for a paradigm shift in how AI is applied to quantum materials. It posits that for complex, mechanism-dependent phenomena like MITs, AI should not act as a black-box predictor of properties but as a "mechanism-aware proposal engine." By encoding physical hypotheses and device constraints as controllable generative biases, the framework aims to bridge the gap between the vast space of stable inorganic structures and the rare, specific subspace of switchable electronic matter. The authors advocate for a closed-loop ecosystem where generative models, physics-based simulations, and experimental feedback iteratively refine the search for credible MIT hypotheses, moving the field from simple property optimization to the design of functional phase-transition landscapes.