Early Detectionand Recovery of SCF Convergence Failuresin Automated Quantum Chemistry Workflows via Time-Series Learning

Abstract Self-consistent field (SCF) convergence failures represent a major bottleneck in high-throughput workflows, particularly for open-shell systems and material discovery. To address this challenge, we introduce an autonomous pipeline that treats the electronic structure descriptors from early spin-unrestricted SCF iterations as a time-series problem. Using statistical descriptors extracted from only ten iterations, a lightweight gradient-boosted classifier accurately predicts convergence failures. The model achieves 94.8% accuracy on 55,000 doublet anionic QM9 test molecules after being trained on only 10,000 examples. We then couple this early warning predictor with restart-based level-shifting interventions derived from Bayesian optimization of the α- and β-shifting parameters. Across six diverse SCF test sets, our pipeline successfully rescued 53.9% of inherently difficult-to-converge molecules while maintaining a 96.3% convergence rate for easy-to-converge cases, reducing the total number of iterations across the 1,200 calculations by 248,355 steps. Overall, this proof-of-concept demonstrates that treating early SCF iterations as a time-series problem provides a data-efficient active-monitoring layer, which can be paired with various remedies to fine-tune convergence parameters and enable more autonomous, large-scale electronic structure workflows.