Agentic multi-fidelity learning of quasiparticle and excitonic properties

Researchers introduce an agent-guided multi-fidelity machine learning framework that corrects numerical instabilities in GW-Bethe-Salpeter calculations for simulating electronic and optical properties of strained MoS2-WS2 bilayers. The approach uses confidence-weighted structural agents and Gaussian process corrections to improve accuracy of quasiparticle gaps and exciton binding energies while preserving physical strain dependence.

This research addresses a fundamental challenge in computational materials science: the numerical fragility of many-body excited-state calculations that are essential for designing optoelectronic devices. GW-Bethe-Salpeter equation methods provide accurate predictions of electronic structure and optical properties but are computationally expensive and prone to convergence failures that escape detection in automated workflows. The proposed agent-guided framework represents a methodological shift toward diagnostic-first surrogate modeling rather than direct interpolation of raw first-principles data.

The work stems from broader trends in computational materials discovery where high-throughput screening demands both accuracy and reliability. Traditional machine learning approaches treat numerical errors as noise to be smoothed away, but this study demonstrates that explicit diagnosis of failure modes—spike-like excursions, near-zero-gap collapse, and dielectric screening artifacts—is essential for trustworthy predictions. The structural agent assigns confidence weights that prioritize high-accuracy reference points strategically, dramatically reducing computational overhead.

For materials researchers and device engineers, this framework enables faster discovery cycles for optoelectronic nanomaterials without sacrificing confidence in predictions. The approach scales beyond the demonstrated MoS2-WS2 bilayers to quantum dots, nanoribbons, and perovskites, domains where accurate optical property prediction directly drives technological applications. Industries developing next-generation solar cells, LEDs, and quantum technologies benefit from accelerated validation pipelines.

Future adoption hinges on whether the method generalizes across diverse material families and computational codes. The framework's emphasis on failure diagnosis over brute-force data aggregation suggests a maturing field moving toward principled uncertainty quantification in materials informatics.

• →Agent-guided multi-fidelity framework corrects numerical instabilities in excited-state calculations without erasing physical phenomena.
• →Explicit diagnosis of computational fragility enables trustworthy surrogate models that outperform direct first-principles data interpolation.
• →Framework applies to quantum dots, nanoribbons, 2D semiconductors, and perovskites facing strong quantum confinement.
• →Confidence weighting and selective high-accuracy reference points reduce computational cost while improving prediction reliability.
• →Methodology addresses critical bottleneck in high-throughput optoelectronic materials discovery pipelines.