A self-evolving agent for explainable diagnosis of DFT-experiment band-gap mismatch

This paper introduces XDFT, a self-evolving closed-loop agent that automatically diagnoses and resolves the common mismatch between DFT-predicted metallic states and experimental semiconducting behaviors in complex materials by iteratively testing curated hypotheses and updating a Bayesian posterior, achieving a 78% success rate on a benchmark of 124 materials.

The Gist

Imagine you are a detective trying to solve a mystery: Why does a computer program keep getting the wrong answer about a material's "personality"?

In the world of materials science, there is a standard computer program called DFT (Density Functional Theory). It's like a very fast, very popular weather forecaster. For most materials, it predicts the weather (electronic behavior) perfectly. But for a specific group of tricky materials, the forecaster keeps saying, "It's going to be a metallic storm!" (conducting electricity like a wire), while the actual experiment shows, "No, it's actually a sunny day with a clear gap!" (acting as a semiconductor).

For years, scientists had to manually check these mismatches one by one, guessing what went wrong. It was slow and tedious.

This paper introduces XDFT, a new "self-evolving detective agent" that automates this diagnosis. Here is how it works, using simple analogies:

1. The Detective's Toolkit (The Hypothesis Library)

Imagine XDFT has a massive toolbox containing 41 different "fixes" or theories. These aren't just random guesses; they are specific scientific adjustments, such as:

• "Maybe the atoms are arranged in a different shape?" (Polymorph)
• "Maybe there's a missing atom or an extra one?" (Defect)
• "Maybe the electrons are acting weirdly and need a special rule?" (Hubbard correction/Magnetism)

2. The Closed-Loop Loop (The Investigation)

Instead of trying every tool at once, XDFT acts like a smart detective who learns as it goes:

1. Pick: It looks at a material and picks the most likely tool from its box to try first.
2. Apply: It runs a complex computer simulation (the "experiment") using that tool.
3. Judge: It compares the result to the real-world experiment. Did the material finally act like a semiconductor?
   • Yes: Great! It records the win and moves to the next material.
   • No: It marks that tool as "less likely to work for this type of case" and tries a different tool.
4. Learn: This is the "self-evolving" part. Every time it solves a case, it updates its global memory. If it learns that "Polymorph" fixes work great for one type of material, it becomes more likely to try that tool first for the next similar material. It gets smarter with every case it solves.

3. The Results: A Detective's Success Story

The team tested XDFT on 124 materials known to be tricky.

• The Problem: 90 of these materials had the "wrong personality" mismatch.
• The Old Way: If you guessed randomly, you'd only get it right about 19% of the time. If you used a standard AI (LLM) without learning, it was only 20%.
• The XDFT Way: XDFT solved 78% of the mismatches (70 out of 90).
• Efficiency: It didn't just get more right; it got there faster. On average, it found the answer in 2.7 tries instead of 4.3, saving a huge amount of computer power.

4. The "Aha!" Moment: A Hidden Pattern

After solving 70 cases, XDFT didn't just give a list of answers; it revealed a hidden pattern, like a detective realizing, "Oh! All the red cars have flat tires, and all the blue cars have broken engines."

The agent discovered a simple rule based on the type of element in the material:

• Main-group elements: Usually need a different shape (Polymorph).
• Transition metals (d-block): Usually need a magnetic fix (Magnetism + U).
• Rare-earth metals (f-block): Usually need a magnetic fix (Bare Magnetism).

The team turned this into a simple four-line rule that anyone can use without needing the complex AI agent.

5. What About the Cases It Couldn't Solve?

XDFT hit a wall on 20 materials. But even this was useful. The paper explains that these failures weren't random; they pointed to specific, very complex physics (like "intermediate valence" or "multiplet structures") that the current toolbox simply doesn't have a tool for yet.

• The Value: Instead of just failing, XDFT acts as a report card, telling scientists exactly what new tools they need to build for the next version of the software.

Summary

XDFT is a self-teaching detective. It doesn't just run calculations; it diagnoses why the standard calculations fail. It learns from every success to get faster and smarter, turning a messy, manual guessing game into a streamlined, explainable process. It successfully fixed the "personality mismatch" for nearly 80% of the tricky materials it tested and provided a clear map of what physics is still missing from our current tools.

Technical Summary

1. Problem Statement

Standard Density Functional Theory (DFT), particularly in its commodity form (GGA-PBE exchange-correlation with perfect crystals and default magnetic configurations), frequently fails to predict the correct electronic ground state for correlated and structurally complex materials.

• The Mismatch: DFT routinely predicts metallic behavior for materials that experiments confirm are semiconductors (finite band gaps).
• The Cause: This mismatch encodes specific "non-idealities" excluded by default protocols, such as strong electron correlation, specific magnetic ordering, alternative polymorphs, or defects.
• The Limitation: Current infrastructure automates execution but lacks the ability to automatically diagnose why a mismatch occurs. Existing workflows either rely on manual expert intervention or apply static, pre-tabulated corrections (like $GGA+U$) only to known chemistries, failing to scale or provide mechanistic explanations.

2. Methodology: The XDFT Agent

The authors introduce XDFT (eXplainable DFT), a closed-loop autonomous agent designed to diagnose mismatches and identify the physical mechanism responsible.

Architecture

XDFT operates as a four-module closed loop:

1. Sherlock (Hypothesis Selection): Maintains a global, self-evolving Bayesian posterior over a library of 41 hypothesis classes (e.g., point defects, magnetic ordering, Hubbard $U$, polymorphs, strain, vdW corrections, spin-orbit coupling). It selects the next hypothesis to test based on the current posterior, an LLM advisor's score, and element-class filters.
2. Manipulator (Atom-Edit): Translates the selected hypothesis into concrete structural or input-file modifications using a tool catalogue (e.g., generating defect supercells, swapping cations, applying strain, setting Hubbard parameters).
3. Simulation: Executes the modified first-principles calculations using VASP on GPU hardware.
4. Comparator (Verdict): Evaluates the resulting electronic character against experimental data. It returns a categorical verdict: Match (gap matches character), Partial (gap exists but magnitude is off), or Fail.

Learning Mechanism

• Bayesian Updating: Sherlock updates a Beta($\alpha, \beta$) posterior distribution for each hypothesis based on the verdicts. This posterior is global, meaning outcomes from one material update the prior probabilities for all subsequent materials.
• Self-Evolution: The agent learns in two ways:
  1. Re-ordering: It dynamically adjusts the priority of hypotheses within the existing library.
  2. Prior Adjustment: It adjusts class-level priors as data accumulates (e.g., downgrading the probability of "magnetism" if it consistently fails, or upgrading "polymorphs" if they succeed).
• Termination: The process iterates up to 10 rounds per material. If a match is found, the material is resolved with a mechanistic attribution. If the budget is exhausted, the case is flagged for experimental re-examination or library expansion.

3. Key Contributions

• Autonomous Diagnosis: Shifts the paradigm from "What does DFT predict?" to "Which excluded non-ideality explains the mismatch?"
• Explainability: Unlike "black-box" machine learning models, XDFT provides a mechanistic attribution for every resolution (e.g., "The gap opens because of a specific magnetic ordering").
• Self-Evolving Posterior: Demonstrates a system that calibrates its internal belief state in real-time, aligning its internal confidence with empirical success rates.
• Derivation of Static Rules: The agent's dynamic reasoning converges on a simple, interpretable, four-line static rule based on element class, which can be applied without the agent.

4. Results

The system was benchmarked on 124 candidate materials (98 valid algorithmic runs, 90 of which were confirmed mismatches).

• Resolution Rate: XDFT successfully identified a resolving mechanism for 70 out of 90 mismatch cases (78%).
  • Comparison: This significantly outperforms a uniform-random baseline (19%), a static LLM ordering (20%), and a naive expert ordering (69%).
• Efficiency: The adaptive ordering reduced the average rounds-to-solution from 4.3 to 2.7, effectively halving the computational cost per resolved material.
• Convergence:
  • The internal Bayesian posterior showed strong correlation with empirical success rates (Spearman $\rho$ rising from -0.32 to +0.69 as the benchmark progressed).
  • The agent learned that "polymorph" hypotheses were underutilized initially and upgraded their priority, while "magnetism" hypotheses were downgraded.
• Emergent Taxonomy: Resolved cases clustered into a clear element-class taxonomy:
  • Main-group: Dominantly resolved by alternative polymorphs (83%).
  • d-block: Dominantly resolved by magnetism + Hubbard $U$ (70%).
  • f-block: Dominantly resolved by bare magnetic ordering (57%) or magnet+$U$ (36%).
• Failure Analysis: 20 cases remained unresolved. Analysis revealed these were due to physics not yet in the library (e.g., intermediate valence, multi-channel correlation, charge-density waves), providing a clear roadmap for future library expansion.

5. Significance and Outlook

• Scientific Impact: XDFT demonstrates that autonomous agents can move beyond high-throughput execution to autonomous diagnosis, extracting interpretable physical insights from systematic theory-experiment disagreements.
• Practical Utility: The derived static rule (Main-group $\to$ Polymorph; d-block $\to$ Magnet+$U$; f-block $\to$ Bare Magnet) allows researchers to apply these insights immediately without running the full agent.
• Scalability: The framework is extensible to other properties (carrier type, magnetic ground state, effective mass) and can incorporate dynamic hypothesis generation (using LLMs to draft new physics classes) for future versions.
• Limitations: The current benchmark focuses only on metal/semiconductor character mismatches, not quantitative gap magnitude errors. The hypothesis library was static during this study, though the architecture supports dynamic expansion.

In summary, XDFT represents a significant step toward self-correcting computational materials science, where agents not only calculate properties but also autonomously identify and correct the underlying physical assumptions when theory diverges from reality.