Tag
The paper presents Conservative Hybrid Graph Network (CHGN), a method for modeling dynamic process systems that allows extrapolation to unseen larger graphs without retraining and improves fault prediction in industrial applications.
The paper resolves contradictions in using neural networks for derivative-free optimization by identifying three key factors—role, radius, and room—that determine when learned local models improve performance in simulation optimization.
HI-MGN introduces a hierarchical multiscale graph neural network to improve long-range communication in mesh-based physics simulations, enhancing accuracy while reducing training time and memory usage compared to existing methods.
This paper introduces agentic Bayesian optimization, where an LLM agent acts as the central decision-maker in the BO loop with a Bayesian backend, enabling online strategy revision and problem reframing. The authors instantiate this in Sara and lenz, demonstrating reliability and performance gains over standard BO and LLM-based baselines.
Introduces Physics-Audited Agentic SciML (PA-SciML), a verification-first workflow where LLM agents discover surrogate models and validate them against physics requirements such as boundary conditions and causality, not just error metrics. Numerical examples show improved trustworthiness over error-only baselines.
Introduces TRIE, an evaluation framework for stochastic PDE surrogates that tests reproduction of invariant measures, trustworthy predictive uncertainty, and efficiency. Benchmarks pointwise-trained neural surrogates, approximate uncertainty methods, and generative models on two SPDEs, finding generative models most consistent.
This paper presents a generalized adaptive sequential sampling method for constructing polynomial chaos expansion surrogate models to improve uncertainty quantification in multi-output engineering structures, balancing variance contribution and spatial exploration.
This paper introduces breakeven complexity, a metric to determine when neural PDE solvers become cost-effective compared to traditional numerical solvers. The framework uses scaling laws to allocate training budgets and evaluates multiple neural solvers on diverse PDE benchmarks.