Epistemic Closure: Autonomous Mechanism Completion for Physically Consistent Simulation

This paper introduces a Neuro-Symbolic Generative Agent that overcomes the "Implicit Context" problem in scientific discovery by autonomously validating and completing physical mechanisms through dimensionless scaling analysis, thereby preventing physical hallucinations and ensuring thermodynamically consistent simulations.

The Gist

Imagine you are trying to build a complex model of a volcano using a very smart, but slightly naive, robot assistant. You give the robot a stack of old textbooks and say, "Build me a simulation of what happens when this volcano heats up."

The robot reads the books, finds a famous equation about heat and pressure, and starts coding. But here's the problem: The textbooks have hidden assumptions. One book says, "Assume the rock is sealed tight like a pressure cooker." Another says, "Assume the rock is porous like a sponge."

If you just ask a standard AI to "write code based on the books," it might blindly follow the "pressure cooker" rule. It will calculate that the pressure gets so high the rock explodes. But in reality, your specific rock is a sponge, the water can escape, and nothing explodes. The AI has created a "Physical Hallucination": the code works perfectly, but the physics is wrong.

This paper introduces a new kind of AI agent called "Epistemic Closure" that solves this problem. Here is how it works, using simple analogies:

1. The Problem: The "Blind Translator" vs. The "Cognitive Supervisor"

• The Old Way (The Blind Translator): Imagine a translator who knows every word in a dictionary but doesn't understand the story. If you ask them to translate a recipe for a cake, but you are actually baking a soufflé, they might still write down the cake instructions because that's what the book said. They don't realize the context is wrong. In science, this leads to simulations that look real but predict disasters that never happen.
• The New Way (The Cognitive Supervisor): This new AI is like a senior engineer sitting next to the robot. It doesn't just translate words into code; it thinks about the physics. It asks: "Wait, the book says 'sealed,' but our rock is hot and permeable. If we seal it, the pressure will blow up. But if we let it breathe, it will be fine. Let's fix the recipe."

2. How the AI Thinks: "Constitutive Skills" and "The Toolbox"

The researchers taught the AI to break down scientific laws into "Constitutive Skills."

• Think of these skills as Lego bricks or tools in a toolbox.
• One brick is "Thermal Expansion" (heat makes things grow).
• One brick is "Darcy Flow" (water moving through a sponge).
• One brick is "Capillary Action" (water sticking to glass).

Usually, an AI just grabs the first brick it finds in the book. This new AI acts like a master builder. It looks at the specific job (the simulation) and asks:

• "Do I need the 'Capillary' brick?" -> Checks the conditions: "No, the rock is already soaked with water. That brick is useless right now." -> Prunes (removes) it.
• "Do I need the 'Darcy Flow' brick?" -> Checks the conditions: "The book didn't mention it, but the rock is porous and the heat is rising fast. If I don't add this, the pressure will explode." -> Completes (adds) it.

3. The Secret Weapon: The "Deborah Number" (The Speedometer)

How does the AI know when to add the "Darcy Flow" brick? It uses a concept called the Deborah Number.

Imagine you are pouring honey into a cup.

• If you pour it slowly, the honey has time to flow out the bottom. (This is a "Drained" state).
• If you pour it instantly, the honey has no time to flow, and it piles up. (This is an "Undrained" state).

The AI acts like a speedometer. It calculates: "Is the heat coming in faster than the water can escape?"

• In this specific experiment (heating sandstone), the AI calculated that the water escapes 100 times faster than the heat arrives.
• The AI realized: "The book assumed the water was trapped (Undrained), but our speedometer says the water is escaping easily (Drained)."
• So, the AI autonomously corrected the physics, adding the missing "escape route" for the water.

4. The Result: Saving the Simulation

• The Naive Model (The Blind Translator): Followed the book blindly. It predicted the rock would explode and fracture because the pressure got too high. This was a lie.
• The New Agent (The Cognitive Supervisor): Realized the water could escape. It added the missing physics. It predicted the pressure would stabilize, and the rock would remain safe. This matched reality.

Why This Matters

This paper shows that AI is evolving from a typist (who just writes code based on what it reads) into a scientist (who understands why the code works).

It proves that AI can:

1. Read scientific papers.
2. Spot hidden assumptions that don't fit the current situation.
3. Fix the math by adding missing pieces of physics that the original author might have taken for granted.

In short, this AI doesn't just write the script for the movie; it checks the script to make sure the physics makes sense before the cameras start rolling. It prevents the "Physical Hallucinations" that could lead to dangerous engineering errors in the real world.

Technical Summary

Here is a detailed technical summary of the paper "Epistemic Closure: Autonomous Mechanism Completion for Physically Consistent Simulation."

1. Problem Statement: The "Implicit Context" and Physical Hallucination

The integration of Large Language Models (LLMs) into scientific discovery is currently hindered by the "Implicit Context" problem. Scientific literature often presents governing equations bounded by invisible assumptions (e.g., "undrained conditions," "isothermal," or "small strain") that are obvious to human experts but opaque to standard generative models.

When LLMs attempt to synthesize simulation code based on retrieved literature, they often fall into "Physical Hallucination": generating code that is syntactically correct and mathematically solvable but physically invalid because it applies a law outside its domain of validity.

• Current Limitation: Traditional numerical solvers (FEM/FVM) act as "agnostic calculators." They lack the cognitive agency to question input validity. If fed a model with incorrect assumptions (e.g., a rigid "undrained" simplification for a drained system), they faithfully converge to erroneous, physically disastrous solutions.
• The Gap: There is a missing "Cognitive Supervisor" capable of autonomously validating, pruning, and completing physical mechanisms before numerical execution.

2. Methodology: Neuro-Symbolic Generative Agent

The authors propose a Neuro-Symbolic Generative Agent that functions as a cognitive supervisor atop traditional numerical engines. The framework operates through a three-stage autonomous workflow:

A. Semantic Encapsulation (Constitutive Skills)

Instead of treating equations as static text, the Agent converts them into modular "Constitutive Skills" ($S$). Each skill is a structured schema containing:

• Functional Form ($F$): The symbolic governing equation.
• Interaction Manifold ($\Omega_{int}$): Input/output physical fields defining connectivity.
• Applicability Constraints ($C_{meta}$): Semantic metadata (e.g., saturation states, material symmetry).

B. Physics-Informed Cognitive Reasoning

The Agent employs a Chain-of-Thought (CoT) workflow to construct a causal graph of the simulation:

1. Deductive Pruning: The Agent evaluates retrieved skills against initial conditions. If a mechanism is mathematically redundant (e.g., capillary pressure in a fully saturated state, $S_r=1$), it is autonomously pruned.
2. Inductive Completion via Dimensionless Scaling: The Agent performs Dimensionless Scaling Analysis (specifically calculating the Deborah number, $De = \tau_{diff} / \tau_{load}$) to determine the dominant physical regime.
   • If the literature implies an assumption (e.g., "undrained") that contradicts the calculated regime (e.g., $De \ll 1$ implies "drained"), the Agent detects a mechanism deficit.
   • It then activates Intrinsic Priors (latent physical knowledge within the LLM, such as Darcy's Law) to inject missing dissipation terms ($\nabla \cdot v$) into the governing equations.

C. Automated Variational Synthesis

Once the constitutive logic is finalized, the Agent translates the abstract physical graph into a weak-form variational formulation and compiles it into executable Python code (using FEniCSx) for finite element simulation.

3. Key Contributions

• Epistemic Correction: The paper demonstrates that AI can move beyond being a "stochastic parrot" (text completion) to acting as an "epistemic partner" capable of challenging and correcting theoretical assumptions embedded in training data.
• Autonomous Mechanism Completion: The system can autonomously identify missing physics (e.g., fluid dissipation in a drained regime) that standard literature retrieval misses, effectively "hallucinating" in a constructive, physically constrained manner.
• Neuro-Symbolic Integration: The framework successfully bridges unstructured scientific literature (RAG) with rigorous numerical execution by using the LLM's latent space as a repository of universal conservation laws (Intrinsic Priors).

4. Results: Thermal Pressurization in Rothbach Sandstone

The framework was validated on a challenging geomechanical problem: thermal pressurization in low-permeability sandstone (Rothbach sandstone).

• The Scenario: A cylindrical specimen is subjected to rapid heating ($200^\circ$C boundary).
• Baseline (Literature-Only): A naive model retrieved from literature assumed "undrained" conditions ($De \gg 1$).
  • Result: It predicted a catastrophic accumulation of pore pressure, causing the effective stress path to cross the failure envelope (Tension Cutoff/Hvorslev Surface), falsely predicting rock fracture.
• Agent-Completed Model:
  • Reasoning: The Agent calculated $De \approx 10^{-2}$ to $10^{-3}$, identifying the system operates in a **Drained Regime** ($De \ll 1$).
  • Action: It autonomously activated the Darcy flow mechanism (Intrinsic Prior) to allow fluid dissipation.
  • Result: The simulation predicted a stable stress path ($p' \approx 8.9$ MPa) that remained well within the elastic envelope, consistent with experimental reality.
• Spatial Verification: The Agent's model correctly reproduced spatial heterogeneity, showing a pressure gradient driving fluid from the center to the drainage boundary, whereas the naive model showed unphysical uniform pressure accumulation.

5. Significance and Future Outlook

• Paradigm Shift: This work shifts the role of AI in science from a "coding assistant" to a "reasoning partner" that safeguards simulations against implicit cognitive biases.
• Safety Valves: The study highlights that Intrinsic Priors (fundamental conservation laws stored in the LLM) act as safety valves, preventing physical violations even when retrieved literature is incomplete or contextually inappropriate.
• Generalizability: While demonstrated on Thermo-Hydro-Mechanical (THM) coupling, the logic (Skill Encapsulation $\to$ Topological Pruning $\to$ Intrinsic Completion) is agnostic to specific physics and could be applied to chemo-mechanics or thermo-magnetism.
• Limitations & Future Work: The system currently relies on the quality of the "seed literature" for functional forms. Future iterations aim to integrate Data-Driven Discovery (e.g., symbolic regression) to allow the Agent to correct the kernel functions themselves, not just the topology.

Conclusion: The paper establishes a new paradigm where AI agents achieve Epistemic Closure, ensuring that scientific simulations are not just mathematically consistent but physically valid by autonomously bridging the gap between theoretical assumptions and physical reality.