Skip to content

Production Surrogate Modelling Auto-Research Multi-Agentic System

Nov 2026 · ADIPEC · 0 citations · 5 references

Abstract

Building surrogate models for production system simulators is a recurring engineering task that remains manual, expert-dependent, and inconsistent across teams. Each update cycle requires a specialist to select input parameters, run simulations, choose ML architectures, tune hyperparameters, validate results, and package the model, spanning days to weeks. Conventional AutoML partially automates model selection but wastes resources on architectures more complex than necessary. This paper presents a multi-agent framework, powered by large language models (LLMs), that automates the full surrogate modelling lifecycle for multiphase production systems in two operating modes: conversational (expert-in-the-loop) and autonomous (self-optimizing until convergence). The framework makes four contributions. First, an autonomous Dataset Generation pipeline, which consists of a three-agent DAG (Sampling Designer, Simulation Runner, Data Reporter) that infers the relevant input parameters, their physically bounded ranges, and an appropriate Latin Hypercube sample size directly from a natural-language description of the well, dispatches the resulting design to the simulator, and returns an audited, reported dataset, removing the manual setup and scripting that otherwise precede every development cycle. Second, a complexity-aware experiment planning strategy in which every autonomous round starts from a default roster of fast, classical regressors spanning a linear, a kernel, and an ensemble-tree family, and escalates to deep-learning or gradient-boosted architectures only through the refinement loop, when a round's best result has failed to meet the stopping criterion. This eliminates the principal inefficiency of conventional AutoML, which is unnecessary evaluation of heavy models. Third, a self-optimizing autonomous mode in which the agent evaluates results against a negotiated stopping criterion, a minimum test R2 and a maximum iteration budget, and, if unsatisfied, autonomously formulates a refinement strategy. This includes adjusting search spaces, substituting architectures, modifying feature engineering, or escalating from classical ML to deep learning, continuing until convergence or an iteration budget is exhausted. Fourth, a DAG-orchestrated pipeline of eight specialized sub-agents across three coordinated DAGs (Sampling Designer, Simulation Runner, Data Reporter, Data Analyzer, ML Planner, Coder, Trainer, Reporter) that produces self-contained training scripts, cross-validated model artifacts, comparison reports, and a plug-and-play Predictor module encapsulating the full preprocessing and inference pipeline. The framework is demonstrated on gas-lift production curve prediction, using data its own Data Generation sub-agent produced from a steady-state production simulator hosted on a dedicated virtual machine, with five input features and 20-element multi-output targets. In autonomous mode, the default three-candidate roster converges within a single round: a random-forest surrogate reaches test-set R2 of 0.98, above the stopping criterion, so no refinement round and no escalation to more expensive candidates are needed. Execution – script generation, training, and report generation across all candidates – completes in under a minute once the plan is confirmed. A structural-constraint-preserving feedback mechanism ensures data paths, output dimensionality, and feature definitions remain consistent across all replanning and refinement cycles. Results show that the framework reduces surrogate model development time from days to minutes, eliminates team-to-team inconsistency, and produces standardized, reproducible artifacts with a deployment-ready predictor for integration into production optimization and digital twin workflows.

View source

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.