Agentic Surrogates: Automating Proxy Models of Simulators with Compute Aware Intelligence

ICLR 2026 Conference Submission21306 Authors

Keywords: Agentic Surrogates, Science Simulators, Adaptive Ensemble Model, Proxy Models, Sample Efficient Surrogates

TL;DR: We introduce an generative AI agentic, compute-aware framework that automates surrogate model construction, adaptively reasoning & orchestrating ensembles and acquisition policies to deliver faster, more efficient, and principled simulation proxies

Abstract:

Proxy (surrogate) models are indispensable for accelerating scientific computation, yet creating them remains a manual, sample-inefficient, and non-reproducible process—especially when simulators are costly and constrained by physics. We present a domain-agnostic agentic framework that automates end-to-end surrogate construction for high-fidelity simulators. At its core, a reasoning engine orchestrates the entire loop, encompassing space-filling candidate generation, batched simulator invocation, warm-start retraining, uncertainty/calibration, and dynamic acquisition switching to achieve user-specified accuracy with minimal wall-clock time and simulator calls. Crucially, this reasoning engine also intelligently constructs and orchestrates an adaptive ensemble of models, where individual components or techniques are dynamically selected based on the specific characteristics of different input and output combinations. For instance, if the agent discerns simpler dependencies, such as linear relationships or reliance on fewer inputs within a particular output ensemble, it can deploy models with reduced complexity or optimized for computational efficiency. Conversely, for complex, highly non-linear, or high-dimensional problems, it will automatically integrate and leverage more sophisticated architectures within this ensemble. This approach ensures that the overall surrogate is a finely tuned composition of expert models, each optimally suited to distinct aspects of the simulator's behavior. The controller manages acquisition rules as a portfolio of experts (residual-error, MC-Dropout variance, EI/EGO, hybrid, randomized) and optimizes a compute-aware objective (error reduction per minute), with extensible support for multi-fidelity scheduling. The framework remains neutral to underlying architecture (e.g., ANNs, PINNs, operator networks such as FNO/DeepONet) and integrates physics-aware stopping alongside MC-Dropout and conformal prediction for calibrated uncertainty.
We evaluate this approach on energy-related scientific modeling problems, including industrial-style flow/process simulators and a PDE proxy. Our findings reveal consistent and substantial performance improvements compared to fixed strategies, demonstrating notable advancements in both sample efficiency and time-to-target accuracy. Prior pilots in oil & gas demonstrated substantially reduced predictive errors compared to the best single acquisition policy while converging faster; here we generalize the method beyond any single domain, preserving these gains under diverse physics and cost profiles. A key novelty lies in the central contribution of introducing compute-aware regret guarantees, which are emphasized up front.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 21306