Epistemic Closure: Autonomous Mechanism Completion for Physically Consistent Simulation

TL;DR

This work establishes a paradigm where AI agents transcend the role of coding assistants to act as epistemic partners, capable of reasoning about and correcting the theoretical assumptions embedded in scientific data.

Abstract

The integration of Large Language Models (LLMs) into scientific discovery is currently hindered by the Implicit Context problem, where governing equations extracted from literature contain invisible thermodynamic assumptions (e.g., undrained conditions) that standard generative models fail to recognize. This leads to Physical Hallucination: the generation of syntactically correct solvers that faithfully execute physically invalid laws. Here, we introduce a Neuro-Symbolic Generative Agent that functions as a cognitive supervisor atop traditional numerical engines. By encapsulating physical laws into modular Constitutive Skills and leveraging latent intrinsic priors, the Agent employs a Chain-of-Thought reasoning workflow to autonomously validate, prune, and complete physical mechanisms. We demonstrate this capability on the challenge of thermal pressurization in low-permeability sandstone. While a standard literature-retrieval baseline erroneously predicts catastrophic material failure by blindly adopting a rigid "undrained" simplification, our Agent autonomously identifies the system as operating in a drained regime (Deborah number De << 1) via dimensionless scaling analysis. Consequently, it inductively completes the missing dissipation mechanism (Darcy flow) required to satisfy boundary constraints, predicting a stable stress path consistent with experimental reality. This work establishes a paradigm where AI agents transcend the role of coding assistants to act as epistemic partners, capable of reasoning about and correcting the theoretical assumptions embedded in scientific data.

Figures (6)

• Figure 1: Architectural overview of the Neuro-Symbolic Generative Agent.a, From Text to Skills: The Agent ingests unstructured scientific literature (PDFs), extracting governing equations into modular "Constitutive Skills." Unlike a naive LLM (grey path) that blindly executes retrieved code, our Agent treats these skills as candidate hypotheses. b, The Reasoning Core: The Agent acts as a cognitive gatekeeper. It employs Deductive Pruning (Scissors icon) to remove mathematically redundant mechanisms (e.g., capillary pressure in saturated states) and Inductive Completion (Blue path) to inject missing physics. The decision logic is governed by Dimensionless Scaling Analysis (Central Dial), specifically using the Deborah number ($De$) to arbitrate between competing mechanisms. c, Physical Consequence: A comparison of simulation outcomes for heated saturated rock. The "Literature-Only" model (top), constrained by the extracted "undrained" assumption, predicts false material fracture (Physical Hallucination). In contrast, the "Agent-Completed" model (bottom) correctly identifies the drained regime, activating Darcy flow to relieve pressure and predicting a stable, physically valid state.
• Figure 2: Visualization of the implicit causal topology utilized by the Agent for autonomous mechanism construction. This graph is not an explicit input but an extracted representation of the Agent's underlying cognitive process, fusing Intrinsic Priors (LLM-based) with Retrieved Skills (literature-based). Dark Blue Path (Retrieved): Represents the primary thermodynamic driver implicitly anchored to the Skill Library (Source II), linking thermal expansion to pore pressure. Green Path (Pruned): Represents the saturation-capillary constraint (Source I). The Agent internally identifies this mechanism as valid but dynamically executes pruning due to the specific simulation context ($S_r=1$), filtering out mathematical redundancy. Orange Path (Intrinsic): Represents a mechanism completion pathway derived from the LLM’s internal physical knowledge rather than external text. It reflects the Agent's latent capacity for fluid dissipation (Darcy flow), preserved to ensure physical stability against time-scale violations (discussed in Section \ref{['sec2.3']}).
• Figure 3: Element-wise verification of the literature-retrieved constitutive kernels.a, Validation of the capillary dynamics extracted from Source I (The Green Path). The Agent-simulated capillary rise height $h(t)$ (solid blue line) accurately captures the transient rise and converges to the theoretical equilibrium limit defined by Jurin's Law (dashed black line, $H = 29.68$ cm). This confirms the Agent's capability to correctly instantiate the saturation-pressure relationship before it is autonomously pruned. b, Validation of the thermal pressurization driver extracted from Source II (The Dark Blue Path). The simulation accurately reproduces the experimental pore pressure evolution from Ghabezloo & Sulem (2009) (orange markers) under undrained heating conditions. This establishes the accuracy of the primary thermodynamic driving force prior to the activation of the Agent's intrinsic completion mechanism (The Orange Path).
• Figure 4: Autonomous regime detection via Deborah number scaling analysis.a, Evolution of the Deborah number ($De$) versus temperature for various rock permeabilities. The bold yellow line represents the target constitutive model used in this study (Rothbach sandstone, $k=10^{-16}$ m$^2$). The Agent identifies that the system operates entirely within the Pink Zone ($De \ll 1$), signifying a "Drained" regime where fluid diffusion dominates. This scaling insight explicitly contradicts the "undrained" assumption often implied in simplified thermal pressurization models (Source II), compelling the Agent to activate intrinsic dissipation mechanisms. b, Decomposition of the hydraulic diffusivity components relative to their initial values at $T_0$. The Agent tracks the competing effects of viscosity reduction (Orange line, promoting flow) and storage capacity changes (Green line), synthesizing them into a net diffusivity trend (Blue line) to modulate the Darcy flow intensity dynamically.
• Figure 5: Emergent validity demonstrated via effective stress path stability. The plot tracks the evolution of the system in $p'-q$ space relative to the rock's failure envelopes (Red Zones: Tensile/Shear Failure). Dashed Grey Line (Literature-Only): Represents the model constrained strictly to the "undrained" literature extraction. The unchecked thermal pressurization drives the effective stress path into the failure zone (marked by the Red Cross), falsely predicting rock fracture. Solid Blue Line (Agent-Completed): Represents the fully coupled model where the Agent has autonomously activated its intrinsic mechanism (Darcy flow) upon detecting the $De \ll 1$ regime. The diffusive dissipation arrests the stress reduction, stabilizing the path at a safe endpoint ($p' = 8.9$ MPa) well within the elastic envelope. This comparison highlights that the Agent's mechanism completion is essential for preserving the qualitative physical correctness of the simulation.