Sertac Bayhan

Teams working on electric-vehicle (EV) infrastructure often need to translate physical operating criteria into executable routing and charging solvers, yet code synthesis with large language models (LLMs) can produce incorrect or constraint-violating code. We present an evaluation-driven, agentic refinement framework in which a human specifies criteria, benchmark scenarios, and acceptance rules, while an agentic LLM system generates, executes, diagnoses, and repairs solver code. We develop a criterion-tagged benchmark suite that serves as a guide for algorithm development and as a standard for evaluating the generated solvers across four EV benchmarks, covering smart routing and smart charging in synthetic, real-world, and simulator-based scenarios, with a held-out test split reserved exclusively for final evaluation. Furthermore, component ablation studies demonstrate that judge-guided feedback is essential for algorithmic code convergence and that diagnostic tracing further improves its reliability. The full system, incorporating diagnostic tracing and the judge LLM, achieved convergence across all twelve calibration configurations. In contrast, removing diagnostic traces caused three configurations to stall, while one-shot baselines achieved only 0–80% success. On development splits, full-system exact-match success ranged from 96.9–100% for routing and 80.0–84.4% for charging, with remaining mismatches accepted by the judge LLM as functionally equivalent. On held-out test splits, exact-match success ranged from 80–100%, and manual inspection found that 100% of outputs satisfied domain-specific acceptance criteria. These results show that diagnostic, judge-guided iteration supports reliable solver development while preserving human authority over specifications, benchmark design, and acceptance criteria.

AI data centers experience rapid fluctuations in power demand due to the heterogeneity of computational tasks that they have to support. For example, the power profile of inference and training of large language models (LLMs) is quite distinct and big divergences can result in the instability of the electric grid. In this paper we propose, to the best of our knowledge, the first physics-informed DLinear time-series model that can accurately forecast power utilization of an AI data center 5-80 minutes (short-term forecasting) into the future. The physics, based on a multi-node lumped thermal resistance-capacitance (RC) network consistent with Newton's law of cooling, is captured using newly derived time-dependent ordinary differential equations (ODE) that separately models and interlinks power consumption with the GPU compute and memory utilization and temperature. The resulting model, that we refer to as PI-DLinear, trained and evaluated on a real AI data center dataset, is not only more accurate than the state-of-the-art (SOTA) models tested, but the forecast profile respects the underlying physics under power throttling and load transient events. Relative to the SOTA transformer-based and non-transformer-based models, improvements in forecasting accuracy (averaged across all look-back and prediction windows) range from 0.782%--39.08% for MSE, 0.993%--51.82% for MAE, and 0.370%--22.28% for RMSE.

The integration of AI-enabled automation into electric vehicle (EV) infrastructure enables domain ex-perts with limited programming expertise to develop complex charging and routing solvers. We presentan evaluation-driven, human-in-the-loop framework in which domain experts specify high-level physicalcriteria and initial prompts, while a Large Language Model (LLM) agent autonomously generates, tests,and repairs solver code through iterative refinement. An evaluator executes candidate code and providesdetailed failure evidence—including criterion-level mismatches, execution traces, and feedback from an LLM-based judge—to guide prompt updates. We demonstrate this workflow on four industrial benchmarks: twosmart routing tasks (synthetic and real Doha road networks) and two smart charging tasks (synthetic andsimulator-based). Ablation and prompt-sensitivity studies on the calibration set show that both diagnostictraces and iterative refinement are critical for reliable convergence. The complete feedback configurationconsistently converges on all development split benchmarks, achieving up to 100% exact-match successon routing and 80.0–84.4% on charging (with judge-accepted tie-breaks), compared to 0–80% for one-shotbaselines. On the test split, our framework attains 80–100% exact-match success, with 100% of outputsmeeting domain-specific acceptance criteria under manual inspection. These results demonstrate that ex-plainable, criterion-grounded feedback loops support effective human-in-the-loop algorithm development,enabling engineers to retain authority over specifications while delegating code exploration and repair to AIagents.

Physics-Informed Neural Networks have recently emerged as a promising approach for applying deep neural networks to solve partial differential equations. However, accurately addressing challenging regions in the solutions of stiff partial differential equations necessitates adaptive methods. Additionally, the inherent limitations of baseline Physics-Informed Neural Networks in handling sequential or time-series data significantly constrain their applicability. In light of this, this paper introduces a Self-Adaptive Physics-Informed Attention-based Gated Recurrent Unit model, which enhances the baseline Physics-Informed Neural Networks framework to address these critical issues. The proposed model represents a significant advancement in Physics-Informed Neural Networks by integrating an attention-based Gated Recurrent Unit layer, thereby enabling more effective modeling of sequential data than existing physics-informed models. Moreover, this paper jointly minimizes the prediction error between the estimated and measured outputs (data-fit loss) and the residual associated with the governing degradation physics (physics-residual loss), while learning adaptive weights to balance both objectives during training. The efficacy of the proposed model is demonstrated through an essential case study, which is to predict the state of health of lithium-ion batteries, utilizing four battery datasets from different sources, namely the NASA, MIT, XJTU, and HUST battery datasets. The proposed model is compared with the baseline Physics-Informed Neural Networks and other benchmark models to demonstrate improvements in predictive accuracy and network initialization capabilities.

The growing occurrence and severity of natural disasters worldwide have heightened the likelihood of major blackouts for electricity customers, as these events significantly impact transmission and distribution systems. Therefore, equipping the distribution network with restoration capability becomes even more important as it can decrease the duration and intensity of the blackouts. In this regard, the rapid growth of grid-forming inverters (GFMI) along with battery energy storage systems (BESSs) creates new opportunities to enhance the distribution network’s ability to shorten blackout duration and intensity for the electricity customers. However, optimal utilization of this high number of GFMIs in the restoration process, considering grid and communication uncertainties caused by severe weather conditions, is a challenging task. This article examines the major challenges and presents a research roadmap to enable effective restoration of active distribution networks dominated by GFMIs. The complexities associated with various stages of the restoration process, the impact of uncertainties, and the dynamic challenges arising from the low-inertia nature of modern power systems are investigated through several case studies using a realistically modified IEEE 123-bus system.