Agentic multi-fidelity learning of quasiparticle and excitonic properties

This paper introduces an agent-guided multi-fidelity learning framework that employs a structural agent to diagnose numerical instabilities in GW-Bethe-Salpeter calculations and applies machine learning corrections to accurately predict quasiparticle and excitonic properties in strained MoS2-WS2 bilayers, demonstrating that explicit detection of numerical fragility is essential for reliable surrogate modeling of excited-state materials.

The Gist

Imagine you are trying to map the terrain of a new, mysterious island. You want to know exactly where the mountains are, where the valleys lie, and how the landscape changes as you walk from one side to the other.

In the world of computer science and materials, this "island" is a new type of ultra-thin material (specifically, a sandwich of two different crystals: Molybdenum Disulfide and Tungsten Disulfide). Scientists want to predict how this material behaves when you stretch or squeeze it (strain), because that changes how it conducts electricity and handles light.

To get this map, they use a super-powerful but very finicky tool called GW-BSE. Think of this tool as a high-tech drone that flies over the island to take measurements.

The Problem: The Drone Gets Confused

The problem is that this drone is incredibly expensive to run and sometimes gets "glitchy."

• The Glitch: Sometimes, when the drone flies over a specific spot (a specific way the crystals are stacked or a specific amount of stretch), it suddenly screams, "There's a mountain here!" when there is actually a flat plain. Or it says, "The ground is zero feet high!" when it should be solid.
• The Cause: These glitches happen because the drone's sensors get confused by a specific type of atmospheric interference (called "long-wavelength dielectric screening"). It's not that the island changed; it's that the drone's math broke down for a split second.
• The Danger: If you just take all the drone's photos and feed them into a computer program to learn the map, the computer will learn the glitches as if they were real mountains. It will think the island is full of fake spikes and holes.

The Solution: The "Agent" Detective

The authors of this paper introduced a new system to fix this. They call it an Agentic Multi-Fidelity Framework. Here is how it works in simple terms:

1. The Multi-Fidelity Drone Fleet: Instead of just one drone, they send out a fleet. Some drones are "low-fidelity" (fast, cheap, but a bit blurry). Some are "high-fidelity" (slow, expensive, but crystal clear). They fly over the same spots to see if they agree.
2. The Agent (The Detective): Before the computer tries to learn the map, a smart "Agent" (a specialized AI assistant) reviews every single photo the drones took.
   • The Agent looks for "spikes" (sudden, weird jumps in the data).
   • It checks if the blurry drone and the clear drone agree.
   • It looks for "near-zero" errors that shouldn't exist.
3. The Verdict: The Agent doesn't just delete the bad photos. Instead, it assigns a "Trust Score" to each one.
   • "This photo is perfect. Trust it 100%."
   • "This photo looks a bit shaky. Trust it 50%."
   • "This photo is clearly broken. Ignore it for learning, but keep it in the back pocket just in case."

The Learning Process: Drawing the Map

Once the Agent has sorted the photos, the computer (using a method called Machine Learning) draws the final map.

• It uses the "low-fidelity" photos to get the general shape of the island (the big trends).
• It uses the "high-fidelity" photos to pin down the exact details.
• Crucially, because the Agent told the computer to ignore the "glitchy" photos, the computer doesn't learn the fake mountains. It learns the real physics of how the material stretches.

The Result: A Reliable Map with a "Confidence Meter"

The final output isn't just a map; it's a map with a Confidence Meter.

• In areas where the data was smooth and the drones agreed, the map is very precise, and the confidence meter is high.
• In areas where the drones struggled or the math was tricky, the map still shows the best guess, but the confidence meter flashes yellow, saying, "We aren't 100% sure here yet."

Why This Matters

The paper shows that you can't just run expensive computer simulations and hope the results are perfect. Sometimes, the computer makes subtle mistakes that look like real science.

By adding this "Agent Detective" layer, they can take a messy, glitchy pile of data and turn it into a clean, reliable guide. This allows scientists to design better materials for electronics and solar cells without wasting time chasing fake data errors.

In short: They built a system where a smart AI detective filters out the computer's math errors before a learning program tries to understand the material, ensuring the final map is accurate and trustworthy.

Technical Summary

Technical Summary: Agentic Multi-Fidelity Learning of Quasiparticle and Excitonic Properties

Problem Statement
Accurate simulation of electronic structure and optical properties in low-dimensional nanomaterials, such as strained van der Waals heterobilayers, relies on many-body $GW$–Bethe–Salpeter equation (GW–BSE) calculations. While physically powerful, these methods are computationally expensive and prone to localized numerical instabilities. Specifically, in quasi-two-dimensional systems, the long-wavelength dielectric screening is fragile; incomplete reciprocal-space convergence can lead to "silent" failures where calculations complete successfully but produce numerically unstable outputs, such as spike-like excursions, near-zero-gap collapses, or cross-mesh inconsistencies.

In high-throughput workflows, treating all completed data points as equally reliable training labels for machine learning (ML) is problematic. If a flexible surrogate model is trained directly on raw first-principles data containing these structured numerical artifacts, it risks learning the numerical pathology alongside the physical trends, thereby degrading predictive accuracy. Current multi-fidelity learning approaches can transfer information across theory levels but often fail to distinguish between physically meaningful trends and numerically contaminated data points.

Methodology
The authors propose an agent-guided multi-fidelity framework designed to correct GW–BSE excited-state landscapes for strained MoS$_2$-WS$_2$ bilayers. The workflow operates as a closed loop involving four distinct stages:

1. Multi-Fidelity Data Generation:

   • Calculations are performed across four stacking registries, two strain branches (in-plane biaxial and interlayer-modulating $c$-strain), and three reciprocal-space samplings ($k$-meshes: $12\times12\times1$, $15\times15\times1$, and $21\times21\times1$).
   • The workflow proceeds through ground-state DFT relaxation, exact diagonalization, $GW$ quasiparticle correction, and BSE exciton calculation.

2. Structured Agent for Quality Control (Triage):

   • A structured agent (implemented as a GPT-5.5 API call) evaluates the raw dataset using precomputed numerical diagnostics.
   • Input Diagnostics: The agent receives workflow completion status, spike metrics (based on normalized second finite differences along the strain axis), cross-mesh disagreement, near-zero-gap indicators, symmetry consistency, and numerical metadata.
   • Output Decisions: The agent assigns a trust score, anomaly class, likely cause, recommended action, anchor priority, and training weight to each data point.
   • Mechanism: Points exhibiting spike-like excursions, near-zero-gap behavior, or strong cross-fidelity disagreement are downweighted or excluded. High-value, sparse high-fidelity anchors are prioritized. This step converts diagnostic signals into decisions about how strongly each point influences learning.

3. Multi-Fidelity Transfer Learning:

   • Graph-Attention Transfer: A graph-attention model learns strain-ordered relationships across environments and reciprocal-space fidelities. It is trained in "delta space" to learn how targets change between different $k$-mesh fidelities (e.g., from $12\times12\times1$ to $15\times15\times1$), rather than fitting high-fidelity targets from scratch.
   • Gaussian-Process (GP) Residual Correction: A GP model is fitted on the residuals (difference between the graph prediction and the target) using a compact set of physically meaningful parent variables identified via causal-structure analysis. This stage removes remaining structured errors and provides calibrated predictive uncertainty.

4. Soft Guardrails:

   • For $GW$ targets in fragile extreme-strain regimes, a soft guardrail prevents unphysical gap collapse by raising predictions to a floor defined by local neighbors and exact-diagonalization references if specific fragility criteria are met.

Key Results

• Identification of Numerical Fragility: The study confirms that anomalies in the GW–BSE landscapes (spikes, near-zero gaps) are linked to incomplete convergence of the dielectric head in the small-$|G|$ regime (long-wavelength screening). These are structured failures, not random noise.
• Agent-Guided Improvement: Parity plots comparing models trained with and without the agent show that the agent-guided approach significantly reduces systematic bias and tightens the distribution of errors. Without the agent, the surrogate learns numerical contamination; with the agent, it recovers physically plausible strain dependencies.
• Reconstruction Accuracy: The framework successfully reconstructs corrected GW direct bandgaps and exciton binding energies. In regions with raw data anomalies, the corrected trajectories suppress isolated artifacts while preserving well-supported slope changes and crossover structures.
• Uncertainty Calibration: The predictive uncertainty estimates are empirically calibrated, closely following ideal Gaussian expectations. Uncertainty is naturally higher in regions of numerical fragility (e.g., near crossover regimes or where cross-mesh disagreement is high) and lower in smooth, high-trust regions.

Significance and Claims
The paper claims that reliable surrogate learning for excited-state materials requires explicit diagnosis of numerical fragility rather than the direct interpolation of raw first-principles data. The primary contribution is a workflow that:

1. Distinguishes reliable training evidence from numerically fragile outputs in large, stage-wise GW–BSE datasets that are impractical to inspect manually.
2. Uses a curated dataset to reconstruct expensive properties via transfer learning, correcting localized artifacts while preserving physical strain dependence.
3. Demonstrates that a closed loop of simulation, diagnostic evaluation, agent-guided data selection, and multi-fidelity learning is necessary for adaptive excited-state computation.

The authors position this framework as a template for other optoelectronic nanomaterials (quantum dots, nanoribbons, 2D semiconductors, hybrid perovskites) where expensive first-principles calculations may fail through subtle convergence errors. They emphasize that the framework mitigates the consequences of numerical fragility rather than removing the underlying convergence difficulty, and that its generalizability to other material families requires further testing.