WorksComputer Science
Reliability by Architecture: Agentic Workflows for Validated Multi-Physics Simulation from Natural Language
Abstract
Preparing, executing, and validating strongly coupled multi-physics simulations require expert time. Agentic workflows use large language models (LLMs) to coordinate them from natural-language descriptions, but reliability requires checks of physical consistency and fidelity to the problem description. Our agentic harness implements explicit task allocation: LLMs interpret user requests and simulator diagnostics, generate input components not covered by predefined construction rules, and propose repairs, while deterministic routines construct supported components and check inputs, edits, and results. We evaluate CO2-storage problems using the multi-physics simulator GEOS, from compositional flow to flow-poromechanics with frictional fault contact. In 80 attempts on two flow problems, grounding raises the proportion of structurally valid inputs from 5/20 to 20/20, but acceptance reaches 20/20 only with complete rule-based input preparation. With identical inputs, local repairs correct three seeded defects while preserving valid inputs. Supporting tests identify simulations passing physics checks despite differing from the problem description or reference results.
Introduction
Problem setting
In geologic storage, multi-physics simulations of coupled subsurface flow and geomechanics inform regulatory assessments [26] and analyses of injection-induced deformation, as at In Salah [20].
A multi-physics simulation requires setting up physical models, geometry, material properties, initial and boundary conditions where applicable, and numerical settings. These choices may be expressed, for example, in configuration files, as in the open-source simulators GEOS for subsurface engineering applications [6, 22] and OpenFOAM for computational fluid dynamics (CFD) [18]. Material and numerical parameters must be consistent, the mesh must be compatible with the selected discretization and appropriately represent any physical interfaces present, and constitutive models, initial and boundary conditions, and coupled solver settings must be mutually consistent.
Once the setup is assembled, running the simulation can require repeated attempts, with manual diagnosis and correction of failures. Diagnostics from parsing, initialization, or solution may not identify the modeling choice responsible. Even if the simulation executes successfully, the computed solution may not satisfy all relevant physical constraints; a completed calculation may also represent a different physical model, initial state, or operating control from the one requested.
Checking the setup and simulated response for physical consistency and agreement with the intended experiment therefore requires expertise in both the physical models and the simulator’s input and solver requirements. These labor-intensive workflows place high demands on valuable expert time. Agentic frameworks, which combine large language models (LLMs) with external tool use [30, 28], offer a way to alleviate this burden by assisting with simulation setup, execution, and validation.
Prior approaches and remaining challenges
LLM agents combine language-model reasoning with retrieval, external tools, and stateful orchestration [30, 21, 27, 28, 11]. An agent uses an LLM to interpret a task and select actions or tool calls. LLM-assisted simulation workflows have applied these capabilities to CFD, porous-media flow, reactor thermal hydraulics, atomistic simulation, and mechanics and geomechanics [4, 14, 1, 25, 19, 9, 2]. Across these studies, documentation and examples guide input generation, static checks assess generated code, and simulator execution provides diagnostics for repair and validation. Outside simulation-specific workflows, LLM refinement methods use self-critiques or reflections on previous attempts to guide subsequent responses [15, 24], while software-agent methods use executable tests to assess repairs [10, 29].
We identify four recurring failure modes that arise when LLMs generate or revise simulation inputs. First, the LLM can generate input components that the simulator cannot accept. Retrieval of documentation, examples, or templates can reduce obsolete syntax and missing-file errors [13, 4, 1]. However, supplying this context does not enforce the input requirements: an LLM-written component may still use unsupported syntax, omit required structure, or refer to an undefined object.
Second, the LLM can combine individually valid components into a physically inconsistent setup or one that differs from the request. Initial conditions, constitutive models, and operating controls may be incompatible; an admissible value or material may be assigned to the wrong region, or a schedule may differ from the requested operation. These errors concern the relationships between components and their intended use. Documentation and structural checks can establish which input constructs are permitted without verifying all relationships between them or agreement with the request.
Third, an LLM-proposed correction can disrupt parts of the input that were already working. Execution-grounded simulation workflows use simulator diagnostics to revise generated inputs [25, 14]. If those corrections are not constrained, however, they may modify valid parts of the input, remove a requested feature to avoid the failure, or repair an edit that has not resolved the error.
Fourth, errors in a setup generated or repaired by the LLM can remain undetected even when the simulation completes. JUTULGPT [14] uses simulator-enforced physics checks and verifies selected user requirements in porous-media flow. Strongly-coupled problems combining flow and geomechanics place greater demands on validation: checks must address consistency across interacting processes and constitutive models, and fidelity to more complex problem descriptions.
These failure modes motivate an explicit allocation of tasks between LLM reasoning and deterministic code. Hybrid methods delegate numerical subproblems to external solvers [9], while SimuCost reports that LLM-based parameter tuning can be slower than conventional scans in its main comparison across 11 simulators, with an additional production code assessed separately [3].
This motivates distinguishing operations that benefit from LLM reasoning from construction, calculation, and checking procedures that can be implemented in deterministic code.
Our contribution
We develop an agentic harness that coordinates LLM interpretation and generation, deterministic input construction, simulator execution, checks, and controlled repair. Deterministic routines construct input components covered by implemented construction rules and check the assembled inputs and simulation results. The LLM interprets problem descriptions and simulator diagnostics, generates input components not constructed by deterministic routines, and proposes edits when no predefined correction applies. The agentic harness constrains LLM-proposed repairs to limit changes to working inputs and requires successful execution, applicable physics checks, and agreement with the requested experiment before accepting a simulation.
We use GEOS, an open-source simulator for coupled subsurface flow and geomechanics [22], as the simulation engine. GEOS has been used in real-world case studies, such as at the Newberry enhanced geothermal system demonstration site, where it coupled thermal, hydraulic, and mechanical processes to predict microseismic activity for comparison with field observations [23]. Other field-based applications include modeling -injection-induced fault opening at In Salah [5] and hydraulic-fracture growth at the Hydraulic Fracturing Test Site 1 (HFTS-1) in the Permian Basin [8]. GEOS’s combination of fluid transport, rock deformation, and fault mechanics capabilities makes this simulator a suitable testbed for our agentic harness, which we demonstrate on -storage examples ranging from compositional flow, which tracks fluid components across phases, to coupled flow–poromechanics with frictional fault contact.
We assess input construction, coupled simulation capability, and repair through distinct evaluations. An 80-attempt comparison examines how grounding and deterministic preparation affect the fraction of accepted simulations. A separate contact demonstration examines the construction, execution, and validation of a coupled setup. Repair is evaluated through paired runs starting from identical copies of an input, with repair disabled in one run and enabled in the other. These comparisons assess whether local repairs correct specified configuration defects while preserving valid content. Supporting tests examine errors that remain undetected by successful execution or conservation checks alone. Together, these evaluations assess how task allocation and explicit acceptance requirements affect the reliability of LLM-assisted simulation.
Methodology
Workflow implementation
The agentic harness orchestrates simulation setup, execution, validation, and repair around GEOS. Deterministic routines in the agentic harness prepare and correct inputs, invoke GEOS, and check inputs and outputs; GEOS performs the numerical solution of the governing equations. We implement the agentic harness by converting a manual GEOS workflow—preparing the mesh and input deck, running the simulator, and examining the results—into an LLM-assisted workflow orchestrated by LangGraph [11]. GEOS represents the simulation setup through XML input files, referred to here as decks. Their elements and attributes specify the physical and numerical configuration. The GEOS XML schema specifies the allowed input elements and attributes, where they may appear, and their required formats.
LangGraph executes documentation retrieval, mesh generation, deck preparation, execution, and postprocessing as nodes in a workflow graph (Figure 1). Deck preparation includes initial generation when no deck is supplied. Each node may combine LLM calls, deterministic routines, and external tool use. Nodes read and update a shared state containing the problem description, input decks, stage status, and records of execution and validation. The orchestrator node serves as a coordination hub, with attached deterministic routing rules selecting subsequent stages from workflow status, errors, check results, and retry counts. LangGraph executes these transitions and feedback loops. Within the orchestrator node, a locally hosted Llama-3 8B model generates commentary on workflow progress that is retained in the shared state.

Figure 1. Agentic harness coordination and study acceptance. (a) The dotted boundary encloses LangGraph nodes. Gray double-headed arrows summarize coordination and shared-state exchange. Deck preparation includes pre-solve admissibility and input-fidelity checks. Postprocessing includes physics checks and applicable input- and output-fidelity checks. Fidelity checks compare inputs and outputs with the user request. The escalation checkpoint manages pausing and resumption for expert review outside the agentic harness. Red dashed arrows show optional repair after failed postprocessing checks through the orchestrator, including physics and fidelity failures, and runtime-error repair from execution to deck preparation. (b) Study acceptance requires satisfying all listed conditions, including any additional fidelity checks performed outside the LangGraph workflow as part of the study. Panel (b) shows assessment logic and outcomes, not additional LangGraph nodes (§ 2.2.3, § 2.4.1 and § 2.4.2).
Input checks, including input-fidelity checks, operate during deck preparation. The postprocessing node applies physics checks and applicable input- and output-fidelity checks. Additional fidelity checks are performed by separate evaluation scripts outside the LangGraph workflow. The escalation checkpoint pauses the workflow for unresolved input or runtime issues. Expert review takes place outside the agentic harness; the checkpoint enables resumption after intervention. § 2.4 describes the checks and final acceptance. Table 1 summarizes the allocation of tasks between LLM reasoning and deterministic routines in the agentic harness.
| Stage | LLM contribution | Deterministic routines |
| Retrieve and interpret | Use repository browsing and search tools to find reference decks when embedding retrieval is unavailable or returns weak matches | Retrieve decks by embedding similarity; rank candidates against the requested physics; assemble current-schema context |
| Mesh and deck construction | Choose geometry strategy and tools; write components not synthesized by deterministic code | Size the mesh; map regions and materials; convert units; correct links between deck objects; construct input components from requested parameters using predefined rules |
| Physical setup | Extract requested parameters | Calculate initial states and controls; make referenced input files available to the simulator; check admissibility and input fidelity |
| Runtime repair | Diagnose the failure from the supplied log excerpts and propose a local edit when no prescribed correction applies | Select error messages and surrounding log lines for LLM diagnosis; apply specified corrections; check schema compliance and preservation of required, previously valid deck content; enforce retry limits |
| Execution and postprocessing | Optionally generate a narrative summary of the simulation results and the outcomes of the post-solve checks | Invoke GEOS; extract execution information and numerical quantities to check completion and apply physics, input-fidelity, and output-fidelity checks |
Table 1. Allocation of responsibilities between LLMs and deterministic routines in the agentic harness surrounding GEOS. The simulator numerically solves the governing equations.
This sequence implements an interpret–act–validate loop, as used by JutulGPT [14] and related reasoning-and-action methods [30]. Appendix § 15.§ 1 describes the shared state and node interfaces.
Constructing and checking the physical setup
The generated-deck path starts from an LLM-generated GEOS input deck; the supplied-deck path starts from a copy of an existing input. Before execution, deterministic routines prepare the deck for the mesh and requested physical setup, then check it. We call this stage deterministic preparation. Its input checks also apply after repair changes a deck.
LLM interpretation and grounded deck generation
The GEOS input schema evolves between releases. In the schema snapshot used in this study, our parser counts roughly 275 element types and nearly 3,000 attributes. Such an evolving schema can lead to plausible-looking input decks that use obsolete syntax.
In the generated-deck path, the LLM extracts requested geometry, phases, controls, and simulation time. Deterministic code builds components for which explicit constructors are available, and the LLM writes the remaining XML. The supplied-deck path skips initial generation and operates on a copy of the supplied input.
Grounding supplies the deck-generation LLM with current GEOS element documentation, retrieved reference decks [13], and fixed construction instructions. The documentation explains valid attributes and their meanings; reference decks show working combinations. The construction instructions encode additional constraints identified during development of the framework (Appendix § 15.§ 2). This context guides generation but does not enforce input requirements.
Deterministic normalization and synthesis
Deterministic preparation combines normalization and synthesis. Normalization corrects missing or inconsistent declarations, object references, and mesh-region assignments in a candidate deck, and converts units where required. Synthesis calculates dependent quantities and constructs input components from requested parameters, such as well completions and timed injection sources. For example, it calculates hydrostatic pressure using the selected fluid model’s density and sets bottom-hole pressure (BHP) for pressure-controlled wells relative to formation pressure at the specified downhole reference elevation. This keeps initial states and well controls consistent with the fluid model without LLM estimates. Appendices § 13.§ 1 and § 15.§ 2 give the deterministic preparation rules and the contact-specific construction and initialization procedure.
When a request leaves a choice unspecified, deterministic preparation applies a predefined default or stops for clarification if it recognizes the ambiguity and no rule applies. Synthesis records each default it supplies, including the component, field, value, and reason. This assumption log does not include LLM-generated values, whose origin is not identified in the returned XML. Unspecified choices may still go unnoticed if neither LLM interpretation nor a preparation rule identifies them (§ 6).
Encoding preparation in deterministic routines supports reproducibility. With the same rules and software versions, identical candidate input decks and parameters undergo the same deterministic transformations. Subsequently, the saved final deck, mesh, companion files, and execution settings can be used to rerun the simulation with an identical setup. The calculations and input modifications that produced this setup can be inspected in the preparation scripts.
Pre-solve admissibility and input fidelity checks
XML schema validation checks whether the assembled deck uses the required elements and allowed attributes. Passing this structural check does not establish physical admissibility or agreement with the request. Pre-solve admissibility checks ask whether the assembled input satisfies simulator and physical constraints. They test both individual bounds, such as positive solid density, and relationships between components, such as compatible saturation parameters and solver-region assignments.
Input fidelity checks ask a different question: does that admissible setup match the request? A plausible permeability assigned to the wrong region, for example, requires a comparison with the intended material assignment rather than a scalar admissibility bound.
Each check runs when its required inputs are available. A detected mismatch prevents acceptance. If an implemented deterministic correction specifies the required change, it is applied to the candidate deck; otherwise, the input and check result are retained for human review. The workflow does not silently replace a requested value with an admissible alternative.
The input-check rows of Table 2 summarize these distinctions. Appendix § 15.§ 5 summarizes the implemented input checks.
| Check | Question | Example |
| Schema validation | Does the deck use the accepted input structure? | Required elements and allowed attributes |
| Pre-solve admissibility | Are values and component relationships physically and numerically admissible? | Positive solid density; compatible solver-region assignments |
| Input fidelity | Does the assembled setup agree with the request? | Requested injection rate and stopping time |
| Post-solve physics gate | Does the computed behavior satisfy the applicable physics checks? | Increase in stored matches the injected mass |
| Output fidelity | Do saved results satisfy the requested output requirements and agree with any specified reference? | inventory after injection stops and requested simulation duration; baseline-trajectory agreement |
Table 2. Questions answered by the validation stages. Input- and output-fidelity checks assess agreement with the request or a specified reference, rather than physical consistency.
Controlled local repair
During framework development, regenerating a whole deck to remove event warnings also removed the event that drove the solver. The simulator exited without taking a single time step. Correcting the reported error had removed the requested calculation. The repair paths therefore use the existing deck as the basis for correction and check the candidate’s required input structure and physical setup. The supplied-deck path restricts each proposal to a local edit.
After a diagnosed failure, the workflow first considers an applicable, fully-specified deterministic correction. Otherwise, the LLM receives the request, current deck, and simulator diagnostics and proposes a local edit. In the supplied-deck path, deterministic code applies the proposal to a copy of the supplied input deck and checks the schema, input constraints, and preservation of required working content before another solve. A rejected candidate leaves the current deck unchanged.
We call this controlled local repair because both the permitted edit scope and the number of correction attempts are limited. The workflow does not repeat an edit when the same failure recurs at the same element or attribute. It may try another permitted edit until reaching the retry limit, after which an unresolved input or runtime failure is sent for human review. Appendix § 15.§ 4 gives the permitted edit forms and retry limits.
Post-solve repair uses feedback from the postprocessing node. When automatic repair is enabled, a FAIL report with diagnostic feedback can route the workflow back to deck preparation to address the failed physics or fidelity checks. This loop permits at most two retries. A WARN or INCOMPLETE verdict withholds acceptance without initiating repair (§ 2.4.1 and appendix § 15.§ 5). Corrections identified during framework development are incorporated into deterministic preparation in later framework configurations when they can be expressed as predefined rules.
Post-solve validation and final acceptance
The problem description specifies the intended physical scenario, physical laws, parameters, and requirements. The agentic harness constructs the simulation, applies physics checks, and uses input and output fidelity checks to assess agreement with that description. A physics gate assesses the computed physical behavior; input- and output-fidelity checks compare the assembled inputs and saved results, respectively, with the request or a specified reference. These roles are summarized in Table 2; study acceptance is distinct from graph coordination, as shown in Figure 1.
Post-solve physics gate
The physics gate checks quantities extracted from the mesh and saved fields for conservation, admissible ranges, expected trends, and consistency between coupled processes. The simulated physics determines which implemented checks apply; for example, mechanics checks are not needed for a flow-only problem.
The check library and the rules for selecting and reporting checks are reused across the flow, poromechanics, contact, and reference-deck cases. Table 13 and appendix § 15.§ 5 summarize the check categories and applicability policy.
| Stage | Category | Representative checks |
| Pre-solve | Admissibility | Permeability and elastic-modulus bounds; relative-permeability endpoint consistency; injector BHP above formation pressure and below the estimated fracture pressure |
| Post-solve | Conservation | Global inventory balance against estimated injection and production; mass monotonicity during injection |
| Post-solve | Field bounds | Positive absolute pressure and fluid density; saturation and component-fraction bounds and sums; porosity and Biot bounds; PVT temperature range and thermal consistency |
| Post-solve | Containment | Free gas in low-permeability seals, or missing seal contrast in a multi-region case containing free gas |
| Post-solve | Coupling response | Injection-zone pressure response considered with mass delivery; sign of process-induced vertical displacement |
| Post-solve | Numerical checks | Run completion and convergence diagnostics; finite stress and fault displacement jumps (legacy fallback: fault-region displacement) |
Table 13. Representative pre-solve admissibility and post-solve physics-gate checks. Applicability depends on the problem profile. Input- and output-fidelity checks assess agreement with the request or a specified reference.
The gate reports PASS only if every applicable check passes. A check that does not apply is recorded as SKIP. Missing data, an error while evaluating a required check, a warning, or a failed check prevents acceptance. Appendix § 15.§ 5 gives the verdict rules.
Output fidelity checks and final acceptance
Passing physics checks does not establish that a simulation matches the user request. For example, in a closed domain, a mass-conservation check compares the increase in stored with the amount injected in the simulation. This check can pass even if the injection rate differs from the user request. Input-fidelity checks compare the encoded rate and time window with the request; output-fidelity checks compare the resulting mass, simulation duration, and saved-output coverage with it.
The postprocessing node applies physics checks and applicable input- and output-fidelity checks. Failures can trigger controlled local repair and rerunning when automatic repair is enabled and retries remain. Additional fidelity checks performed by separate evaluation scripts outside the LANGGRAPH workflow do not trigger workflow repair.
Final acceptance of a simulation result requires completed execution, a physics-gate PASS, and passing all required input- and output-fidelity checks. Reliability here means withholding acceptance when a required check fails or evidence is missing. Saved inputs, simulator outputs, and check reports make this decision traceable. The injection-schedule and trajectory tests illustrate these criteria (§ 4.4). Passing the specified checks does not establish that the tested scenarios cover the practical problem of interest; experts must assess that applicability before drawing scientific conclusions.
Evaluation design
We simulate four physical scenarios with GEOS: an emplaced plume (T1), a two-well flow case (T2), and two flow–poromechanics cases. The continuum fault-zone case (T3) represents faults as finite-width regions within the porous rock and couples fluid pressure to rock deformation. The frictional-contact case (T4) instead represents faults as discrete surfaces, with a contact model describing relative motion across them. Figure 2 shows the simulation domains, and Table 7 provides the scenario specifications. Table 3 identifies which scenarios test setup, coupled capability, and repair.

Figure 2. Simulation domains of the four scenarios simulated in GEOS, as – cross-sections (not to a common scale; all cases place a storage aquifer between seals, Table 7). Three principal evaluation designs reuse these scenarios, as summarized in Table 3. (a) The flow cases share a layered aquifer between caprock and base seals. T1 starts with an emplaced mobile free-gas plume whose upward migration is checked, whereas T2 starts with 100% brine and uses a bottom-hole-pressure-controlled injector/producer pair. The depicted initial plume belongs to T1 and the wells to T2. (b) The coupled poromechanics case (T3) adds two steeply dipping () continuum fault zones that compartmentalize the domain, with source-driven injection. (c) The SPE11C-inspired contact case (T4) [16], on a shallower section with a larger width-to-depth ratio, treats three steep faults as discrete frictional surfaces, using Lagrange multipliers to enforce contact constraints and a friction coefficient of . Injection uses two volumetric SourceFlux regions; the vertical axis is exaggerated for legibility. Colors: sand = storage reservoir/aquifer, blue-gray = seal, gray = fault, orange = wells or sources, blue = emplaced .
| Evaluation | Inputs and comparison | Question |
| Generated-deck flow ablation | Plume (T1) and two-well flow (T2); four configurations, 10 attempts per case and configuration | Which setup components increase accepted yield? |
| Contact capability demonstration | Natural-language description of flow–poromechanics with frictional contact (T4) | Can the workflow construct, execute, and validate this coupled setup? |
| Paired supplied-deck repair | Two-well flow (T2) and continuum flow–poromechanics (T3); identical valid or defective decks with repair disabled and enabled | Can a local correction restore execution while preserving the required setup? |
Table 3. The three principal evaluations. Physical scenarios are reused to answer different questions; generated and supplied inputs are kept distinct.
Table 7. Physical specifications of the four GEOS scenarios and the SPE11B reference comparison. The scenarios are reused across evaluation designs as mapped in Table 3. SPE11B is an external reference comparison (§ 4.4). “Drive” is the mechanism that mobilizes CO; horizons are the simulated physical times. All domains use cells (T4 uses 22,280 tetrahedra).
| Case | Domain (km) | Added physics & structures | Drive | Horizon |
| T1 flow-only | compositional CO–brine, isothermal ; no wells/faults | emplaced free-gas plume () | 365 d | |
| T2 two-well flow | CO–brine, isothermal ; initially 100% brine; injector & producer under bottom-hole-pressure (BHP) control | well-pair sweep | 365 d | |
| T3 +poromech. & faults | poromechanics; two continuum fault zones; regional-constant materials | SourceFlux: 500 million kg over 730 d, then 182.5 d monitoring | 912.5 d | |
| T4 +contact | frictional Lagrange contact on three faults (, cohesion 2 MPa), fixed fracture storage | two SourceFlux regions, 30 d injection + 15 d monitoring | 45 d | |
| SPE11B | thermal compositional flow; layered facies; no wells/faults | two SourceFlux sources, each 0.035 kg/s: years 0–50 and 25–50 | 1000 yr |
Generated-deck flow ablation
The flow ablation successively adds grounding, normalization, and synthesis to ungrounded generation, giving four configurations. Each configuration has 10 plume attempts (T1) and 10 two-well attempts (T2), for 80 attempts in total. Corresponding attempts request the same sampling seed across configurations (Appendix § 15.§ 6). For each configuration, we report how many generated decks satisfy the XML schema and how many attempts produce an accepted simulation without human intervention. The fraction of attempts accepted in this way is termed accepted yield. For both T1 and T2, all four configurations use the same checks and acceptance criteria.
All configurations use Claude Sonnet 4.6 for deck and mesh generation. The ablation varies the grounding context supplied to the deck-generation LLM and the deterministic preparation applied to the generated deck, while keeping mesh generation unchanged. Grounding comprises documentation, reference decks, and construction instructions (§ 2.2.1). These are enabled or disabled as a package; their individual contributions are not isolated.
When deterministic synthesis is enabled, routines calculate the requested initial state and well controls using the selected fluid model. These include the hydrostatic brine background, plume free-gas saturation, and well pressures relative to formation pressure. When synthesis is disabled, the same checks assess the generated values without constructing replacements.
A comparison separate from the generated-deck ablation tests normalization using 20 grounded decks saved from this ablation. Each deck is run with synthesis enabled, once with normalization and once without it, using the same acceptance criteria and with runtime repair disabled (Appendix § 15.7).
Contact capability demonstration
The contact demonstration uses T4, a layered reservoir with three frictional faults whose permeability depends on normal contact loading through a hydraulic-aperture relation. The fluid system is –brine at 333.15 K with a fixed salinity of 2.5 mol per kilogram of brine. The natural-language description requests 30 days of injection through two source regions followed by 15 days of monitoring.
The problem description specifies geometry, material properties, physical laws, initial loading, injection, and output requirements. The agentic harness supplies the GEOS-specific model and solver configuration through its supported contact construction procedure, generates a fresh mesh and deck, and executes and checks the coupled simulation. The contact-case physics-gate and fidelity checks cover the physical fields, contact behavior, and requested injection schedule. Appendix § 13.§ 1 details the deterministic preparation and contact-specific checks.
Paired supplied-deck repair evaluation
The repair evaluation uses the two-well flow case (T2) and the continuum flow–poromechanics case (T3). T2 is a 365-day simulation with a bottom-hole-pressure-controlled injector and producer (Figure 2, panel a). T3 injects 500 million kg over 730 days and then monitors for 182.5 days in a mesh with two finite-width fault zones (Figure 2, panel b).
For each setup, the paired evaluation compares two runs starting from identical copies of a supplied input deck and its companion files, with repair disabled in one run and enabled in the other. Using supplied inputs keeps generation errors out of the comparison.
We first test each valid baseline, then introduce the same specified defect into both copies and repeat the comparison. Both runs receive the same deterministic preparation and acceptance checks; only the repair-enabled run may receive a checked local correction after a failure. Together, these pairs test whether controlled local repair can correct known configuration defects without changing valid inputs.
In T2, the tests break a well-control reference or remove a relative-permeability table whose values are stated in the request. In T3, the test breaks a solid-solver reference. These three defects stop execution and are not corrected by the existing deterministic preparation routines. They were selected to test known repair paths, rather than sampled to estimate how often repair succeeds on arbitrary defects. Appendix § 15.8 specifies the permitted edit scope and the checks used to ensure that other input-deck content remains unchanged.
Supporting validation and portability tests
The supporting tests examine what can be missed by a completed solve or an individual check. Table 4 maps each validation question to its test; § 4.4 reports the outcomes, and Appendix § 14 gives the supporting methods and evidence.
| Question | Test | Evidence sought |
| How closely does a coupled solver configuration reproduce an analytical solution? | Single-phase Terzaghi consolidation, with separate mesh and timestep refinement | Pressure, displacement, and settlement errors from direct GEOS runs, without agents |
| Can a mass-balanced result violate the requested schedule? | Injection followed by monitoring | Input schedule and saved inventory compared with the request |
| What can output checks miss? | Seven input defects, with pre-solve checks bypassed | Differences between output detection, input admissibility, and fidelity of material assignment |
| Do the implemented check functions detect their intended violations? | Direct physics-gate check-library test | Detection and false alarms without simulation-file extraction |
| How closely do extracted inventories agree with published reference results? | SPE11B reference comparison | inventory histories compared with the official geos1 submission |
| Does passing physics checks imply agreement with a reference trajectory? | Continuum numerical-parameter test | Physics verdicts compared with baseline-trajectory differences |
| Can a known incompatible request be escalated before execution? | Contact–well compatibility and graph-routing tests | Solver incompatibility distinguished from the workflow’s routing response |
Table 4. Supporting validation tests and the questions they examine. The tests distinguish physics checks, fidelity to the request, and workflow responses to failures.
Separate OpenFOAM tests assess portability of input preparation without using the GEOS-based agentic harness. An optional pre-run LLM critic was evaluated separately and is excluded from the primary claims and accepted-yield comparison. It is disabled in the principal evaluations and omitted from Figure 1. Appendix § 14.§ 6 describes its evaluation and results.
Results
Generated-deck flow ablation
Figure 3 reports results for the plume (T1) and two-well (T2) benchmarks and separates the effects of grounding, synthesis, and normalization into three comparisons. Panel (a) compares ungrounded and grounded generation with normalization and synthesis disabled. Grounding raises generation-stage schema validity from 5/20 to 20/20, but neither configuration produces an accepted simulation (Table 5).
Table 5: Flow-ablation outcomes, with 10 attempts per benchmark and configuration. Rows successively add components. Schema validity is assessed after generation and its allowed internal retries; acceptance additionally requires simulation completion, the physics gate, and input- and output-fidelity checks. Intervals are two-sided 95% Wilson intervals.
| Configuration | Schema-valid (both cases) | Accepted Plume | Accepted Two wells | Combined accepted yield [95% CI] |
| Ungrounded generation | 5/20 | 0/10 | 0/10 | 0/20 [0–16.1%] |
| + grounding | 20/20 | 0/10 | 0/10 | 0/20 [0–16.1%] |
| + normalization | 20/20 | 0/10 | 0/10 | 0/20 [0–16.1%] |
| + synthesis | 20/20 | 10/10 | 10/10 | 20/20 [83.9–100%] |
Table 5.

Figure 3. Effects of grounding, synthesis, and normalization. Each ‘without’ or ‘with’ setting includes 20 runs (10 per flow benchmark); error bars are two-sided 95% Wilson intervals. Panels (a) and (b) report the generated-deck flow ablation. Panel (c) reports a separate comparison that reruns the same 20 decks saved from that ablation. Panel (a) measures schema validity; panels (b) and (c) measure final acceptance. The enabled components are shown below each panel, and runtime repair is disabled.
Panel (b) compares separately generated decks with and without deterministic synthesis, keeping grounding and normalization enabled. Without synthesis, the generated decks fail requirements for the requested initial state or well controls, leaving accepted yield at 0/20. Deterministic synthesis calculates these quantities from the extracted parameters and selected fluid model, raising acceptance to 10/10 for each benchmark and thus 20/20 overall. The two-sided 95% Wilson interval is 83.9–100%. Appendix § 15.§ 6 reports which initial-state and well-control requirements were violated and how many attempts were affected in T1 and T2.
Panel (c) shows a separate normalization comparison using identical saved grounded decks, with deterministic synthesis enabled in both runs. Normalization raises acceptance from 9/20 to 20/20 without turning any passing result into a non-passing one: T1 acceptance rises from 9/10 to 10/10, and T2 acceptance from 0/10 to 10/10. Normalization supplies supporting structure required by the synthesized components: it removes a material-model component defined twice, adds missing regional pressure statistics, and corrects an invalid mesh-region layout. Appendix § 15.7 details the failures and the normalization edits that resolve them.
Contact capability demonstration
The prompt-to-result T4 run generates a new mesh with 22,280 tetrahedra and 922 contact faces and a new input deck, then executes the requested schedule in 10.9 minutes without human intervention, including deterministic preparation and postprocessing. The simulation completes without runtime repair.
All 922 faces are classified as sticking, with no opening or sliding, at each of the 9 saved output times. The contact and hydraulic-aperture checks also pass at each time. The final CO₂ inventory agrees with the requested 60 million kg to within 0.378 kg. Its change from day 30 to day 31 is 5.22 × 10⁻⁸ kg, consistent with injection stopping before monitoring. Together, these results satisfy the physics-gate and input- and output-fidelity requirements.
This demonstration is limited to the observed sticking regime. The fixed-fault-storage approximation and its supporting sensitivity analysis are discussed in Appendix § 13.§ 1.1.
Paired supplied-deck repair evaluation
All 20 flow-ablation attempts with complete deterministic preparation finish their first simulation execution. Because runtime repair is invoked only after a failure, it is not exercised in that ablation. We therefore conduct a separate paired supplied-deck evaluation.
In T2, the LLM restores the reference to an existing well-control block or inserts the missing relative-permeability table using values stated in the request. Only the named attribute or table changes, and both corrected simulations pass the physics gate and all 14 input- and output-fidelity checks. In T3, one attribute edit restores the solid-solver reference. The subsequent simulation passes the physics gate and exactly reproduces the baseline in all 495 comparisons of pressure, stress, and displacement across regions and saved times.
All three defective decks stop execution with repair disabled. With repair enabled, each defect is corrected by one local edit, after which the simulation is accepted. Both valid baseline decks pass in both settings. No repair occurs for the valid baselines, so their input decks remain unchanged (Table 6). Appendices § 15.8 and § 15.9 give the repair protocol and trajectory criteria.
| Input | No repair | Repair enabled | Accepted edit or outcome |
| Valid two-well flow deck | PASS | PASS | No repair edit |
| Broken well-control reference | Non-pass | PASS | One attribute |
| Missing relative-permeability table | Non-pass | PASS | One table added |
| Valid continuum flow–poromechanics deck | PASS | PASS | No repair edit |
| Broken solid-solver reference | Non-pass | PASS | One attribute |
Table 6. Matched supplied-deck repair tests. Each pair starts from identical inputs. PASS requires both the physics-gate checks and all required input- and output-fidelity checks to pass.
Supporting validation tests
Analytical verification of coupled consolidation
A separate, deterministically prepared Terzaghi consolidation test compares coupled pressure and displacement with an independently evaluated analytical solution (Appendix § 14.§ 1). Separate mesh and timestep refinement reduce the largest sampled errors. With 80 cells and a maximum timestep of 0.0625 s, the largest relative errors at the four sampled times are 0.334% for pressure, 1.070% for displacement, and 0.820% for loaded-end settlement. TPFA and hybrid mimetic flow give the same reported errors on this orthogonal mesh. These GEOS runs verify the tested single-phase coupled configuration; they do not evaluate agentic setup or repair.
Fidelity to the injection schedule
During development, an injection-and-monitoring problem exposed a failure that mass conservation alone could not detect. The request specified 2 kg/s for 1460 days followed by monitoring without injection until day 3650. Generated decks instead used 10 or 5 kg/s and continued injection. The simulated increase in stored CO₂ matched the mass injected at those rates, so the mass-balance check passed even though the simulations did not follow the request.
Source synthesis constructs the requested injection period. Input-fidelity checks compare the injection rate and injection period specified in the input deck with the problem description, while output-fidelity checks assess the saved CO inventories against the requested injection and monitoring behavior. Five end-to-end runs all pass these checks and the physics gate, with a median elapsed time of 2.38 minutes (Appendix B.3). These five runs evaluate the implemented response to the injection-rate and timing errors.
Complementary validation checks
A separate seven-defect poromechanics test bypasses pre-solve checks to examine what solver execution and the post-solve physics gate detect without input checks (Table 10). Five defective decks abort or fail the gate. The negative-solid-density and caprock-permeability cases complete and pass because neither changes a checked output enough to violate its criterion at the simulated horizon.
| Seeded defect | Direct solve + post-solve physics gate |
| Negative Young’s modulus | aborts at initialization |
| Poisson ratio 0.95 (inadmissible) | aborts at initialization |
| Solver references an undefined region | aborts at initialization |
| Grain modulus below drained modulus | aborts at initialization |
| Reference porosity 1.8 | execution completes; gate FAIL |
| Caprock at storage-zone permeability (no longer a seal) | execution completes; gate PASS—output matches the baseline run |
| Negative solid density | execution completes; gate PASS—invisible to the fluid-density check |
| Valid baseline deck | execution completes; gate PASS |
Table 10. Seven input defects tested by direct execution and the post-solve physics gate, with pre-solve checks bypassed. A gate PASS here does not imply workflow acceptance.
The solid-density error illustrates why input checks remain necessary: the post-solve density check reads fluid density, whereas the complete workflow rejects negative solid density before execution. In the caprock case, the sealing layer is assigned the storage-zone permeability. This value is admissible in isolation but does not represent the requested sealing-layer properties. Detecting the error therefore requires comparing each region’s assigned permeability with the requested reservoir–caprock configuration.
A separate unit-level test supplies scalar and summary quantities directly to the physics-gate functions. The functions detect all 15 introduced violations, while the unchanged baseline triggers no false alarm among its 14 applicable checks (Table 9). This tests the check functions, not extraction from simulation files.
| Check category | Seeded errors | Caught |
| Admissibility (saturation, pressure, composition, porosity, density) | 6 | 6 |
| Conservation ( mass balance, monotonicity) | 2 | 2 |
| Constitutive (Biot bounds, PVT temperature range, fault slip) | 3 | 3 |
| Coupling (poroelastic uplift sign, injectivity) | 2 | 2 |
| Numerical (solver convergence, stress finiteness) | 2 | 2 |
| Total | 15 | 15 |
Table 9. Direct check-library test: introduced violations and detections by category. The unchanged baseline raises no alarm among its 14 applicable checks.
Reference comparisons
A comparison with the SPE11B CO-storage benchmark [16, 17] assesses CO inventory histories extracted from a rerun of an expert-authored GEOS deck against those in the official geos1 submission. It uses saved outputs outside LangGraph and does not evaluate agentic input generation or repair.
Agreement with the published reference is not uniform across quantities. The extracted mobile CO inventory in box A has a relative trajectory error of 1.6% (Figure 4). The other six relative trajectory errors range from 0.504 to 1.677, and the mean over all seven trajectories is 0.887 (Table 11).
Figure 4. SPE11B reference comparison. (a) Layered facies, reporting boxes A and B, and the two injection locations in the SPE11B domain, shown as an 8.4 km × 1.2 km cross-section. (b) Mobile in box A from the rerun and official geos1 submission. Table 11 also reports mobile in box B, immobile and dissolved in both boxes, and in the seal within box A.
| Quantity | Box A | Box B |
| Mobile CO | 0.016 (1.00) | 0.901 (1.34) |
| Immobile CO | 0.504 (1.10) | 0.700 (1.40) |
| Dissolved CO | 1.113 (2.18) | 1.298 (2.39) |
| CO in seal | 1.677 (3.24) | — |
| Mean rel. (7 QoIs) | 0.887 |
Table 11. SPE11B inventory comparison against the official geos1 submission. Entries are relative trajectory errors; parentheses give the rerun value at 1000 years divided by the reference value at the same time. The mean covers the seven reported quantities.
The continuum flow–poromechanics test compares a modified numerical setup with its baseline. It increases the initial time step from to s and also disables nonlinear line search, which damping of Newton updates. Both repair-disabled and repair-enabled executions complete and pass the physics gate, so neither invokes workflow repair. In both runs, the maximum normalized field-increment difference from the baseline is 8.57%, exceeding the specified 1% reproduction criterion. They therefore fail the output fidelity comparison required to reproduce that baseline (Appendix C.9).
Pre-execution escalation
In the tested GEOS build, combining runtime fault generation with an internal well aborts before the first time step. Removing runtime fault generation eliminates the reported well-subregion error. The alternatives in this setup are to split the fault in the mesh beforehand or represent injection by a volumetric source.
The separate graph-routing test requires both runtime fault generation and the internal well to be retained. The workflow detects the conflict before deck generation and sends the request for human review. This tests recognition and escalation of a known incompatibility, rather than exhaustion of runtime retries (Appendix B.4).
Portability of input preparation to another multi-physics simulator
In a separate analysis using OpenFOAM, we examine whether predefined construction rules help produce complete, runnable simulation inputs for a second simulator. We focus on input preparation and do not use the agentic harness of the GEOS-based framework. OpenFOAM requires a directory of input files rather than a single XML deck.
For laminar and turbulent single-mesh dam-break cases, retrieved reference templates yield 10/10 cases passing structural and phase-fraction-bounds checks. A separate comparison uses prose-only grounding without those templates: adding deterministic synthesis raises the laminar pass rate from 10% to 50% and the turbulent pass rate from 10% to 60%.
We also considered a five-region conjugate-heat-transfer case requiring 56 consistent input files. Reference-case grounding supplies the LLM with the complete working case for the same region layout. Without deterministic synthesis, LLM generation yields no runnable case either without grounding or with reference-case grounding (0/5 each); runs with reference-case grounding produce approximately 43 files. Here deterministic synthesis supplies the missing input files before execution, using templates parameterized by the supplied mesh to construct regional dictionaries, fields, and numerical settings. In the GEOS ablation, it instead corrects malformed input components and constructs the requested initial physical state. With reference-case grounding and synthesis, all five cases run without runtime repair and pass structural and heat-transfer checks: bounded temperature, energy-equation convergence, regional continuity, and a nonzero thermal gradient.
Conclusions
We set out to determine how task allocation between LLM reasoning and deterministic operations can improve reliability in LLM-assisted multi-physics simulation. The four conclusions below show how our proposed task allocation addresses the failure modes identified in our review of prior work.
Grounding and deterministic correction play complementary roles in addressing errors in generated input structure. Grounding guides the LLM toward valid input structure through documentation, reference decks, and construction instructions, raising generation-stage schema validity from 5/20 to 20/20. Deterministic normalization then corrects structural
omissions and inconsistencies that remain in the generated deck. The two therefore act at different stages: grounding informs generation, whereas normalization directly modifies the resulting input. In the saved-deck comparison, enabling normalization alongside synthesis raises acceptance from 9/20 to 20/20 without turning a passing result into a non-passing one.
Schema-valid input still needs construction and checking of a consistent physical setup. Schema-valid inputs do not by themselves establish consistency among the physical models, parameters, and numerical configuration. Deterministic synthesis constructs input components using pre-specified rules and model-based calculations, while separate checks assess physical consistency and agreement with the intended experiment. In the evaluated flow cases, adding synthesis raises accepted yield from 0/20 to 20/20.
Controlled local repair can correct configuration errors while preserving working inputs. Paired runs start from identical input decks with repair enabled or disabled. Repair corrects each of three seeded defects with one local edit that passes deterministic checks, while leaving valid baseline decks unchanged. Repairing the continuum flow–poromechanics deck reproduces the baseline pressure, stress, and displacement histories across five regions and 33 saved times. These results show that targeted LLM-based repair, constrained by deterministic checks, can correct defects in an input deck without regenerating the entire deck.
A reusable physics gate and problem-specific fidelity checks provide complementary validation. Our physics gate applies a common check library across flow, poromechanics, and contact problems, selecting the applicable physics checks for each case. Input- and output-fidelity checks assess agreement with the problem description and are defined for the evaluation at hand. The supporting tests show why both are needed: an incorrect injection schedule can conserve mass, and a simulation can pass the physics gate while failing the specified baseline-trajectory comparison.
The frictional-contact demonstration applies this task allocation to a more complex multi-physics setting. A valid setup requires consistent choices across flow, poromechanics, the requested Coulomb contact law, fault hydraulic behavior, fluid-phase ordering, and mechanical initialization. Deterministic preparation implements these coupled requirements, and the completed simulation passes the contact, inventory, and schedule checks. This extension retains LLM interpretation while adding deterministic procedures for the consistency requirements introduced by the new physics.
Within the taxonomy of Lee and Kim [12], which distinguishes assistive, task-autonomous, workflow-autonomous, and research-autonomous systems, our results support workflow autonomy for user-specified simulations in the tested configurations. The agentic harness coordinates multiple simulation stages: it constructs inputs, executes simulations, checks physical consistency and fidelity to the problem description, and performs constrained repairs without routine human intervention. Experts still define the scientific objectives, implement construction rules and acceptance checks, and assess whether the modeling assumptions and simulated scenarios are appropriate for the practical problem of interest. The contribution is therefore improved reliability of unattended workflows, not independent scientific investigation. Retaining this task allocation when the physics or request changes requires implementing and assessing the corresponding construction rules and checks; our results do not establish autonomous transfer to unfamiliar configurations.
Outlook
Future work has three priorities: improving traceability, broadening the evidence for workflow reliability, and establishing case-specific numerical accuracy.
Improve traceability of inputs, transformations, and assumptions. The workflow retains the request, generated inputs, repair records, simulator logs, and validation reports, but does not link every input value or modification to its origin. Values can come from the request, LLM inference, retrieved examples, deterministic rules, simulator defaults, included files, or companion data. A typed intermediate representation and change record could preserve these links through generation, normalization, repair, execution, and validation. This would allow an auditor to reconstruct how the final simulation setup was assembled, identify assumptions introduced at each stage, and direct uncertain choices to additional checks or human review.
Broaden the evidence for workflow reliability. The two flow-ablation benchmarks informed development of the deterministic routines, and the supplied-deck repair cases were chosen because they matched repair paths already implemented in the framework. Neither group was therefore kept separate from framework development. A future held-out evaluation would instead report outcomes for a predefined set of prompts, physical problems, and defects withheld during framework development, i.e., evaluated only after the deterministic routines, checks, and repair policy were frozen.
Matched comparisons across LLM backends and agentic systems would assess sensitivity to the reasoning component. Complete workflow implementations for OpenFOAM and other multi-physics simulators would assess whether the task allocation remains effective beyond the current GEOS workflow and the separate OpenFOAM input-preparation tests.
Extend numerical verification and calibrate validation tolerances. The analytical test verifies single-phase consolidation, not the multiphase and fault-contact configurations of T3 and T4. Numerical convergence has not yet been established for either coupled demonstration. Their tetrahedral TPFA grids also require assessment of nonorthogonality effects. Mesh and timestep error estimates should be used to set tolerances for differences between computed results and reference solutions.
Acknowledgments
Claude Code (using Opus 4 and Fable 5 models) and Codex (using GPT-5 and GPT-6 models) assisted with development of the deterministic tooling around the LLM agents (normalization and synthesis routines, input checks, physics checks, and associated tests), and with manuscript drafting, restructuring, and language revision, under the direction and review of the lead author. The authors retain full responsibility for the scientific claims, references, and conclusions, including the physics judgment, assessment of execution failures, and final validation.
Claude Code and Codex also assisted with figure-generation code and revisions to layouts, labels, and captions. Schematics represent the described workflow and case specifications; charts and plots use recorded experimental results and simulation/reference data.
Funding
This material is based upon work supported by the Department of Energy under Award Number(s) DE-FE0032793. The authors also acknowledge support by the SUETRI-B Reservoir Simulation Industrial Affiliates Program at Stanford University.
Disclaimer
This report was prepared as an account of work sponsored by an agency of the United States Government. Neither the United States Government nor any agency thereof, nor any of their employees, makes any warranty, express or implied, or assumes any legal liability or responsibility for the accuracy, completeness, or usefulness of any information, apparatus, product, or process disclosed, or represents that its use would not infringe privately owned rights. Reference herein to any specific commercial product, process, or service by trade name, trademark, manufacturer, or otherwise does not necessarily constitute or imply its endorsement, recommendation, or favoring by the United States Government or any agency thereof. The views and opinions of authors expressed herein do not necessarily state or reflect those of the United States Government or any agency thereof.
Author contributions
S.T. designed and implemented the agentic framework, conceived the evaluation studies, conducted the numerical experiments, analyzed and interpreted the results, prepared the figures, and drafted the manuscript. E.D. provided overall research guidance and supervision, and reviewed and revised the manuscript. H.A.T. secured funding and provided technical expertise on the use of the GEOS simulator. All authors approved the final manuscript.
Competing interests
The authors declare no competing interests.
Data availability
The SPE11 benchmark specification and the geos1 community reference series are publicly available from the SPE11 project [16, 17].
Evaluation cases and contact-specific methods
Table 7 gives the physical specifications of T1–T4 and the SPE11B reference comparison. The evaluation designs that use these scenarios are mapped in Table 3. For T4, the procedures for establishing initial pressure and stress, calculating fault permeability, and specifying fracture storage are described in Appendix A.1.
Contact construction and initialization
The contact construction procedure maps the requested physical laws and parameters to a shared –brine fluid model and linked fault-storage, hydraulic-aperture, permeability, and friction models. It selects Phillips brine properties, Span–Wagner density, Fenghour viscosity, and the CO2Solubility flash model. The numerical settings below are supplied by this procedure. The framework snapshot is frozen before the reported prompt-to-result run.
The initial state is brine-saturated. Deterministic preparation constructs the hydrostatic pressure profile by sampling the selected GEOS fluid model in independently constructed vertical columns with 1 m and 0.5 m cell spacing. Both profiles are anchored at 101,325 Pa at the 1200 m domain top and compared.
If is the integrated saturated weight above elevation , the prescribed compressive total-stress magnitudes are 101,325 + vertically, 101,325 + 0.9 along Y, and 101,325 + 0.7 along X, in Pa. Deterministic preparation writes these stresses and the matching boundary tractions as negative values, following the GEOS sign convention. Regional Biot coefficients convert the total stresses to effective stresses.
For a deck containing a contact solver and surface-element region, normalization constructs the supporting fault-generation objects from the mesh-exported fault planes. These include one Rectangle per fault, the ruptureState and isFaceSeparable fields, a SurfaceGenerator, and a fracture-generation event. The conformal interfaces split at before the first solve.
Before injection, the workflow performs two mechanical initialization solves and advances the coupled flow–mechanics model with injection disabled. The resulting fields must satisfy the initialization, conservation, and contact-law checks.
The subsequent 45-day simulation uses one MPI rank with a direct linear solver, nonlinear tolerance , and a maximum one-day time step. A 100 kg flow-residual normalization floor permits an absolute mass residual of kg in smaller-inventory cells; inventory and phase-volume checks are applied separately.
Hydraulic aperture and fixed storage
Hydraulic aperture follows the Barton–Bandis relation , where is the magnitude of the compressive normal contact traction, m, and MPa. Permeability is . Geometric storage aperture remains , with unit surface porosity.
A five-case storage-sensitivity study started from the equilibrated, pre-injection state generated by the reported T4 workflow. The runs retained its mesh, initial fluid and stress states, contact parameters, and hydraulic-aperture and permeability relations. Each run covered 30 days of injection and 15 days of monitoring.
Three cases used fixed fault storage at 0.1, 1, and 10 times the baseline value. A matched fixed-storage control changed the geometric aperture and porosity while preserving the baseline initial pore volume and component inventories. A fifth case started with the same initial pore volume and inventories but allowed fault porosity to vary linearly with pressure. In this variant, fault storage changes through pressure-dependent porosity rather than being calculated from the simulated fault deformation. Comparisons at nine common saved times assessed pressure fields, component inventories, and contact states.
Across the four comparisons with the fixed-storage baseline, the largest relative RMS difference in pressure change was 0.000443%, normalized by the baseline pressure-change RMS within each region. No contact-state classifications changed. Initial fault pore volumes ranged from 0.281 to 28.130 m, compared with approximately m of bulk pore volume. These results support fixed storage over this tested range, with all runs remaining in the sticking regime. They do not test fault opening, sliding, or fault-storage volumes comparable to bulk storage.
Supporting validation evidence
Analytical consolidation verification
We consider one-dimensional Terzaghi consolidation [7] in a saturated column without gravity. The loaded end () is drained; the opposite end () is fixed and impermeable, and lateral strain is constrained. The material has Young’s modulus , Poisson ratio , porosity 0.3, permeability , and fluid viscosity . The fluid is incompressible, and a grain bulk modulus of approximates incompressible grains and a unit Biot coefficient. From an unloaded state with zero pressure and displacement, compressive traction rises to over ; the first solver step is .
In the instantaneous-load, incompressible limit, the independent reference is
The series uses 1,000 terms. We use eight-node hexahedral finite elements for mechanics, coupled to either TPFA or hybrid mimetic flow. The meshes are uniform, with or cells. For the hybrid mimetic method, we use the Beirão da Veiga–Lipnikov–Manzini inner product. Pressure and axial displacement are compared at saved cell centers and nodes, respectively, at using relative discrete errors; settlement is the mean axial displacement at the loaded end.
| Cells | Maximum timestep (s) | Pressure (%) | Displacement (%) | Settlement (%) |
| 40 | 0.25 | 1.052 | 3.315 | 2.673 |
| 80 | 0.25 | 0.993 | 3.089 | 2.387 |
| 80 | 0.0625 | 0.334 | 1.070 | 0.820 |
Table 8. Analytical consolidation errors, maximized over the four sampled times, not over the complete trajectory. Both flow methods give the values shown to the reported precision.
Coverage of input and output checks
The check-library test supplies baseline quantities directly to the physics-gate functions and changes one quantity at a time. It tests whether each function detects that change, without reading simulation files or evaluating fidelity to a problem description.
Table 9 groups the 15 perturbations by changed quantity. They exercise 14 applicable checks because two perturbations target different conditions within the saturation-bounds check: a saturation above one and phase saturations that do not sum to one. A violation is counted as detected if the check returns FAIL or WARN. None of the checks raises an alarm on the unchanged baseline.
A separate test introduces seven input defects, one at a time, into a valid continuum poromechanics deck and runs the solver with pre-solve checks bypassed. This exposes what solver execution and the post-solve physics gate detect on their own (Table 10). Negative solid density requires an input check because the saved density checked by the gate is fluid density. Assigning the storage-zone permeability to the caprock is not detected by a permeability admissibility check; it requires comparing the region’s assigned permeability with the requested sealing-layer properties.
Injection-schedule fidelity
The prompt requests injection at 2 kg/s for 1460 days, followed by monitoring without injection until day 3650. The development failures motivating the test are described in § 4.4.
Deterministic synthesis constructs the timed source. Input fidelity checks verify its rate, component, units, and active interval. At shutoff, the increase in inventory from its initial value must agree with the requested 252,288,000 kg injection within 1%. At every saved monitoring time, the change from the shutoff inventory must also remain below 1% of the requested injected mass.
To represent the requested schedule exactly, the workflow ends a simulation time step when injection stops and saves the inventory both before injection and at shutoff. Later saved inventories cover the monitoring period.
Each of the five runs starts from the same problem description and independently generates its mesh and input deck. The largest relative mismatch at shutoff is , and the largest relative change in inventory during monitoring is . The median wall-clock time for the full workflow, from the problem description through validation, is 142.72 s.
Contact–well compatibility and escalation
A focused GEOS test combines runtime fault generation (SurfaceGenerator) with an internal well (InternalWell). The input deck passes schema validation but stops before its first time step with a well-subregion error. Removing runtime fault generation removes the error, identifying this feature combination as incompatible in the tested setup. An internal well can instead be used with a fault already split into the input mesh, or a SourceFlux can replace the well. A separate LangGraph routing test requests runtime fault generation and an internal well while prohibiting either alternative. The feasibility check sends the request to human review before deck generation. The first test identifies the incompatibility in GEOS; the second verifies the workflow’s pre-execution response.
SPE11B reference comparison
In separate evaluation scripts, we compare inventory histories extracted from saved outputs of an expert-authored GEOS SPE11B deck with those from the official geos1 community submission. This comparison does not evaluate LLM-based deck generation, LangGraph orchestration, or automatic repair. SPE11B was used during framework development and is not treated as a held-out evaluation.
Figure 4 shows the domain and mobile trajectory in box A; Table 11 reports the full comparison. The table gives relative trajectory errors over 0–1000 years and, in parentheses, the rerun value at 1000 years divided by the reference value at the same time.
The rerun trajectories are calculated from saved spatial fields at 5-year intervals, whereas the reference trajectories were calculated within the simulator and interpolated to the same comparison times. Both simulations use the 840 × 120 SPE11B spatial grid.
Held-out evaluation of the optional critic
We evaluated the optional critic separately on 12 matched pairs of valid and deliberately defective inputs, excluding cases used during development. Each defective input contradicted one explicit requirement that the existing deterministic checks did not assess. The defective input and its paired valid input therefore both passed those checks before critic review. These reviews did not require a GEOS run. Before evaluation, we specified the performance required to include the critic in the primary claims and accepted-yield comparison: detection of at least 10 of 12 contradictions and approval of at least 11 of 12 valid inputs without edits or human review, with no edits to clean inputs or to fields outside the identified defect.
The critic identified all 12 contradictions. Only 1 of 12 valid inputs was approved unchanged; the other 11 were preserved and routed to human review (Table 12). Seven of 12 proposed corrections to defective inputs passed the existing input checks. One accepted correction also changed an unrelated statistics-output interval. The critic therefore did not meet the predefined thresholds and was excluded from the primary claims and accepted-yield comparison.
| Outcome | Count | Rate | 95% Wilson CI |
| Target contradiction identified | 12/12 | 100% | 75.8–100% |
| Valid original approved unchanged | 1/12 | 8.3% | 1.5–35.4% |
| Defect correction accepted after checks | 7/12 | 58.3% | — |
| Valid original routed to human review | 11/12 | 91.7% | — |
Table 12. Held-out evaluation of the optional critic on 12 matched pairs of valid and deliberately defective inputs. Sensitivity is target-specific defect detection; baseline specificity is the proportion of valid originals approved unchanged. Intervals are two-sided 95% Wilson intervals.
Framework implementation and evaluation protocols
Shared state and node interfaces
The typed state holds the request, candidate inputs, per-stage status, and inter-node data. Repair records include retry counters, a simulator-independent error category, the structural baseline of the first validated deck, previously attempted fixes, and physics feedback. These records support resumption and prevent repeated unsuccessful edits.
LangGraph passes the shared state between nodes and executes the transitions selected by deterministic routing functions. These functions use stage status, error categories, retry counts, and human-review flags to determine progression, retries, and escalation. Deck-validation and runtime-error loops can return directly to deck preparation; failed postprocessing checks with diagnostic feedback return through the orchestrator when automatic repair is enabled and retries remain. Checkpointing preserves the shared state for workflow resumption. Preparation and validation operate within the existing nodes shown in Figure 1; they are not additional graph nodes. The generated-deck path constructs an initial candidate, while the supplied-deck path starts from an existing input.
Grounding and preparation rules
Grounding uses a fixed, locally stored copy of GEOS’s auto-generated data-structure documentation. Per-element tables list attribute names, types, defaults, and descriptions. The XML schema is a separate input to structural validation. Fixed prompt instructions add cross-element constraints encountered during framework development; they guide generation but do not verify that the generated deck satisfies those constraints.
Normalization reconciles the deck with the mesh and supporting input structure, including unit conversion, material references, and cell-block mappings. Synthesis constructs specified components from extracted parameters. Initial-state and well-control calculations use the selected fluid model. Files referenced by the input deck are copied into the simulation working directory before execution.
For a declared fault-free problem, fixed mesh-selection rules choose a structured hexahedral grid and reject unrequested faults. Contact-specific preparation instead uses conformal fault planes as described in Appendix § 13.§ 1. Admissibility and input fidelity checks assess the assembled setup (Appendix § 15.§ 5).
Synthesis records each supplied default’s component, field, value, and reason. The routines return these records through the problem profile when available, and the deck-generation node stores them in the shared state. Generated XML does not identify whether a value came from the request or LLM inference, so the log cannot record all deck assumptions. Logging a default does not require user approval before applying it.
Model configuration
Deck generation, repair, and meshing use Claude Sonnet 4.6 (claude-sonnet-4-6) through an API gateway. Reference decks are retrieved with a local embedding index using all-MiniLM-L6-v2, without an LLM retrieval step. The orchestrator’s progress commentary is generated by the locally hosted Llama-3 8B model llama3-groq-tool-use:8b, served through Ollama, and retained in the shared state. Workflow routing uses the deterministic rules described in § 2.1.
Both API-hosted and locally hosted LLMs use a sampling temperature of 0.1. The held-out critic study uses Claude Sonnet 4.6 at the same temperature. The flow ablation holds the model choices and sampling temperature fixed across configurations.
Repair implementation and stopping rules
Simulator diagnostics are classified to select a repair route. Deterministic code applies a prescribed correction when one is available; otherwise, the LLM proposes a correction using the existing deck and diagnostics. Checks assess required structure and physical constraints; the preservation mechanisms differ between the supplied- and generated-deck paths.
In the supplied-deck path, a proposal must identify exactly which attribute or XML element it changes. An insertion must specify one existing parent element, and deletion is allowed only for a named element that is not referenced elsewhere. Before the revised deck is used, deterministic checks assess its XML schema, input constraints, physical admissibility, and preservation of required content. An ambiguous or invalid proposal is rejected, leaving the current deck unchanged.
The generated-deck path can instead produce revised XML from the existing deck and diagnostic feedback. Once a baseline is established, a structural-drift check rejects removal of recorded solver or initial-condition element types, required top-level sections, or all well regions. It permits additions and attribute-value changes, so input and post-solve checks remain necessary to assess the corrected setup and results.
The general workflow permits at most 20 correction attempts, reduced to 18 for errors classified as critical or high severity. It can stop earlier when the same error persists after several different corrections or when further changes no longer make progress. The paired supplied-deck evaluation uses a stricter limit of two correction rounds after the initial failed execution (Appendix C.8). Thus, the allowed edit forms limit what may change, and the retry limits bound how many corrections may be attempted.
Repair after failed postprocessing checks has a separate limit of two attempts. Feedback identifies the failed physics or fidelity requirement and its diagnostic quantities. A correction may change the erroneous initial condition, boundary control, or constitutive parameter; the revised deck and simulation must pass the required checks.
Check applicability, tolerances, and verdicts
The problem profile selects the checks applicable to each simulation. Pre-solve admissibility checks assess whether the input satisfies implemented simulator and physical constraints, including bounds and relationships among temperature, elastic moduli, relative-permeability parameters, well pressures, and solver regions. Table 13 summarizes the physics checks by stage; contact and schedule checks are detailed in Appendices A.1 and B.3.
Input fidelity compares the assembled setup with the problem description, while output fidelity compares the saved results with the requested outcome or a specified reference trajectory. One input-fidelity check compares requested permeability with positive values in the deck after converting millidarcies or darcies to square meters. The check passes when at least one deck value is within a factor of 12 of the requested permeability. This loose threshold detects likely copying or unit-conversion errors; it does not verify the regional assignment of every permeability value.
Physics-check tolerances are fixed constants in the implementation and do not vary with mesh or time-step resolution. The trajectory comparisons use separate tolerances defined in Appendix C.9. Future mesh- and time-step-refinement studies are needed to determine how both sets of tolerances should depend on discretization.
An evaluated physics check returns PASS, WARN, or FAIL. An explicitly inapplicable check returns SKIP. Missing fields or an extraction or array-processing error leave an applicable check unevaluated; the aggregate report records INCOMPLETE when no stronger non-passing verdict applies. Acceptance requires all applicable checks to pass and at least one physics check to be applicable.
A FAIL report with diagnostic feedback can initiate post-solve repair when enabled. A WARN or INCOMPLETE verdict withholds acceptance without initiating automatic correction. After the permitted retries, any required check that remains non-passing withholds acceptance. Input- and output-fidelity failures reported by postprocessing can enter this loop; checks performed by separate evaluation scripts outside LANGGRAPH do not trigger workflow repair.
Generated-deck flow ablation
All planned attempts contribute to the reported totals, including those stopped by an input check before simulator execution. Runtime repair is disabled, so a failed attempt is not revised and rerun. Within each benchmark, replicate uses API seed in all four configurations to reduce sampling variation when comparing corresponding attempts across configurations.
Both problem descriptions request a fluid-model-based hydrostatic brine background anchored at 101,325 Pa at elevation 5000 m. T2 additionally requires injector BHP below 1.2 times initial formation pressure and producer BHP below formation pressure.
Input- and output-fidelity checks cover the domain, layers, materials, initial state, 365-day duration, and output schedule. T1 additionally checks plume mobility, upward migration, and closed CO inventory; T2 checks well controls and inventory accounting. T1 does not require the plume to reach the caprock within 365 days. T2 does not assess fracture-pressure safety because the problem description supplies no fracture-pressure threshold.
Without synthesis, the configuration with grounding and normalization fails to reproduce the requested initial pressure and temperature in all 10 plume attempts. In the two-well case, 9 attempts fail the requested well-to-formation pressure relationship; the remaining attempt stops at the pre-solve pressure limit. These counts explain the failures summarized in § 4.1.
Matched normalization comparison
Each of 20 saved grounded decks is processed twice, once with normalization disabled and once with it enabled, giving 40 evaluations. Synthesis and the remaining deterministic preparation are enabled in both runs. Runtime repair is disabled so that any difference can be attributed to normalization.
Each pair uses the same deck, mesh, acceptance criteria, 365-day duration, and 1800-second simulation timeout. All preparation and initialization failures are counted. A preliminary two-deck pilot used to finalize the procedure is not included in the comparison.
All 11 non-passing runs stop before the requested simulation takes a time step. Nine well decks omit the regional-pressure calculation needed to construct their well controls. A tenth well deck also specifies the wrong number of mesh-region entries. The plume deck defines the same constitutive component both independently and within a compound solid model, causing initialization to fail.
Normalization adds the missing pressure calculation, corrects the mesh-region entries, and removes the duplicate constitutive definition. Nine pairs pass in both settings, 11 change from non-pass to pass, and none change from pass to non-pass. The comparison measures normalization as a whole; it does not isolate the contribution of each correction.
Paired supplied-deck repair
Each pair starts from identical copies of an input deck and its referenced files. Runtime repair is disabled in one copy and enabled in the other; all other processing and acceptance requirements are identical. The valid input decks are tested in both settings before the defects are introduced.
After a failure, the repair-enabled run first considers a prescribed deterministic correction. If none applies, the LLM receives the problem description, defective deck, and simulator diagnostics and proposes one local edit. It is not shown the correct result or given instructions for reversing the inserted defect. Before another execution, the deterministic checks in Appendix C.4 must approve the revised deck. Once approved, the revised deck is executed without another LLM call.
The two-well tests break an injector control reference or remove a gas relative-permeability table whose values are in the problem description. Each defect is corrected with one proposal followed by a second execution. Acceptance requires passing the physics gate and all 14 input- and output-fidelity checks. A separate comparison confirms that content outside the approved edit remains unchanged.
The continuum case couples isothermal –brine flow to Biot poroelasticity on 12,822 tetrahedra with two 500 m-thick fault zones. The materials, mesh, loads, and injection location and rate remain fixed. The defect replaces one solid-solver reference with a nonexistent name while retaining the correct solver definition.
Before breaking the solid-solver reference in the continuum flow–poromechanics deck, the valid deck is run once with repair disabled and once with repair enabled. This tests whether enabling repair changes a valid input or its result.
After the solid-solver reference is broken, the repair-enabled run permits at most two correction rounds, allowing no more than three simulation executions including the initial failure. Each execution has a 3600-second limit, and the results are compared using saved VTK fields.
Reference-trajectory comparison for the continuum flow–poromechanics case
The repaired continuum flow–poromechanics simulation and both repair settings of the numerical-parameter test are compared with the same unmodified baseline. All runs use the verified initial state, and pressure, stress, and displacement are saved at 33 specified times, including injection shutoff and the end of monitoring.
Acceptance requires a physics-gate PASS, agreement with the initial state, finite mechanical fields, and no tensile effective stress detected by the stress screen. Inventory error at shutoff and subsequent inventory drift must each remain below 1% of the requested injection. The trajectory comparison uses the change in each field from that simulation’s own initial state, thereby comparing the evolution during injection and monitoring rather than any initial offset.
For each pressure, stress, or displacement field , the normalized trajectory difference is
where is the change from the run’s initial state and is the corresponding change in the unmodified baseline. Cell-field norms are volume weighted; nodal displacement uses the Euclidean norm. Denominator floors are constant fields of 100 Pa for pressure and stress and m for displacement. The 495 comparisons cover five regions, three fields, and 33 times.
Because the repair should restore only the broken reference, its trajectory must reproduce the baseline with . The numerical-parameter test deliberately changes time-stepping and nonlinear-solver settings and uses the less restrictive criterion . These thresholds measure agreement with the baseline; they do not measure spatial or temporal discretization error.
The numerical-parameter test increases initialDt from to s and disables nonlinear line search, retaining 30 Newton iterations, two permitted time-step cuts, and nonlinear tolerance . The saved trajectory is assessed against the baseline even when execution and physics checks pass. The test runs with runtime repair disabled and enabled; neither run invokes repair because both complete and pass the workflow checks. The separate output-fidelity comparison detects their disagreement with the baseline without triggering workflow repair. Its outcome is reported in § 4.4.
References
- [1]Zaid Abulawi, Zavier Ndum Ndum, Eric Cervi, Rui Hu, and Yang Liu. AutoSAM: An agentic framework for automating input file generation for the SAM code with multi-modal retrieval-augmented generation. arXiv preprint arXiv:2603.24736, 2026.arxiv.org/abs/2603.24736
- [2]Yared W. Bekele. GeoSim.AI: AI assistants for numerical simulations in geomechanics. arXiv preprint arXiv:2501.14186, 2025.arxiv.org/abs/2501.14186
- [3]Yadi Cao, Sicheng Lai, Jiahe Huang, Yang Zhang, Zach Lawrence, Rohan Bhakta, Izzy F. Thomas, Mingyun Cao, Chung-Hao Tsai, Zihao Zhou, Yidong Zhao, Hao Liu, Alessandro Marinoni, Alexey Arefiev, and Rose Yu. SimulCost: A cost-aware benchmark and toolkit for automating physics simulations with LLMs. arXiv preprint arXiv:2603.20253, 2026.arxiv.org/abs/2603.20253
- [4]Yuxuan Chen, Xu Zhu, Hua Zhou, and Zhuyin Ren. MetaOpenFOAM: An LLM-based multi-agent framework for CFD. arXiv preprint arXiv:2407.21320, 2024.arxiv.org/abs/2407.21320
- [5]Olga Dufour, Frederic Bourgeois, Cheikh Oumar Ba, Jian Huang, Sylvain Grelaud, and Luc Pauget. CO2 storage risk assessment in large-scale industrial projects: In Salah case study. In Carbon Capture, Utilization, and Storage Conference and Exhibition, pages 161–165, Houston, Texas, 2025.
- [6]GEOS Project. Input files. GEOS documentation, https://geosx-geosx.readthedocs-hosted.com/en/latest/coreComponents/fileIO/doc/InputXMLFiles.html. Accessed October 26, 2026.
- [7]GEOS Project. Poromechanics. GEOS documentation, https://geosx-geosx.readthedocs-hosted.com/en/latest/docs/sphinx/basicExamples/poromechanics/Example.html. Terzaghi-type consolidation example; accessed October 30, 2026.
- [8]Stanislav Glubokovskikh, Christopher S. Sherman, Joseph P. Morris, and David L. Alumbaugh. Transforming microseismic clouds into near real-time visualization of the growing hydraulic fracture. Geophysical Journal International, 234(3):2473–2486, 2023.DOI
- [9]Seokjae Heo. Enhancing reliability and automation of LLM-based structural analysis using a hybrid multi-agent pipeline. Scientific Reports, 16:19458, 2026.DOI
- [10]Carlos E. Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik Narasimhan. SWE-bench: Can language models resolve real-world GitHub issues? In International Conference on Learning Representations (ICLR), 2024. arXiv:2310.06770.arxiv.org/abs/2310.06770
- [11]LangChain. LangGraph. Open-source software repository, https://github.com/langchain-ai/langgraph, 2024. Accessed October 7, 2026.
- [12]Keonhak Lee and Innyoung Kim. Agentic AI for computational fluid dynamics: A review. Applied Mechanics, 7(4):81, 2026.DOI
- [13]Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. Retrieval-augmented generation for knowledge-intensive NLP tasks. In Advances in Neural Information Processing Systems (NeurIPS), 2020. arXiv:2005.11401.arxiv.org/abs/2005.11401
- [14]Knut-Andreas Lie, Olav Møyner, Elling Swee, and Jakob Torben. Agentic scientific simulation: Execution-grounded model construction and reconstruction. arXiv preprint arXiv:2603.00214, 2026.arxiv.org/abs/2603.00214
- [15]Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, Shashank Gupta, Bodhisattwa Prasad Majumder, Katherine Hermann, Sean Welleck, Amir Yazdanbakhsh, and Peter Clark. Self-Refine: Iterative refinement with self-feedback. In Advances in Neural Information Processing Systems (NeurIPS), 2023. arXiv:2303.17651.arxiv.org/abs/2303.17651
- [16]Jan M. Nordbotten, Martin A. Fernø, Bernd Flemisch, Anthony R. Kovscek, and Knut-Andreas Lie. The 11th society of petroleum engineers comparative solution project: Problem definition. SPE Journal, 29(5):2507–2524, 2024.DOI
- [17]Jan M. Nordbotten, Martin A. Fernø, Bernd Flemisch, Anthony R. Kovscek, Knut-Andreas Lie, Jakub W. Both, Olav Møyner, Tor Harald Sandve, Etienne Ahusborde, Sebastian Bauer, et al. Benchmarking CO2 storage simulations: Results from the 11th society of petroleum engineers comparative solution project. International Journal of Greenhouse Gas Control, 148:104519, 2025.
- [18]OpenFOAM. File structure of OpenFOAM cases. OpenFOAM User Guide, Section 2.1, https://www.openfoam.com/documentation/user-guide/2-openfoam-cases/2.1-file-structure-of-openfoam-cases. Accessed October 26, 2026.
- [19]Donggeun Park, Hyeonbin Moon, and Seunghwa Ryu. A self-correcting multi-agent LLM framework for language-based physics simulation and explanation. npj Artificial Intelligence, 2:10, 2026.
- [20]P. S. Ringrose, A. S. Mathieson, I. W. Wright, F. Selama, O. Hansen, R. Bissell, N. Saoula, and J. Midgley. The In Salah CO₂ storage project: Lessons learned and knowledge transfer. Energy Procedia, 37:6226–6236, 2013.DOI
- [21]Timo Schick, Jane Dwivedi-Yu, Roberto Dessi, Roberta Răileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools. In Advances in Neural Information Processing Systems (NeurIPS), 2023. arXiv:2302.04761.arxiv.org/abs/2302.04761
- [22]Randolph R. Settgast, Ryan M. Aronson, Julien R. Besset, et al. GEOS: A performance portable multi-physics simulation framework for subsurface applications. Journal of Open Source Software, 9(102):6973, 2024.DOI
- [23]Christopher S. Sherman and Joseph P. Morris. Modeling microseismic activity in the Newberry enhanced geothermal system. In Proceedings, 42nd Workshop on Geothermal Reservoir Engineering, SGP-TR-212, Stanford, California, 2017.
- [24]Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: Language agents with verbal reinforcement learning. In Advances in Neural Information Processing Systems (NeurIPS), 2023. arXiv:2303.11366.arxiv.org/abs/2303.11366
- [25]Mohammad Soleymanibrojeni, Roland Aydin, Diego Guedes-Sobrinho, Alexandre C. Dias, Maurício J. Piotrowski, Wolfgang Wenzel, and Celso Ricardo Caldeira Rêgo. GENIUS: An agentic AI framework for autonomous design and execution of simulation protocols. Communications Materials, 7:115, 2026. arXiv:2512.06404.DOI
- [26]U.S. Environmental Protection Agency. Federal requirements under the underground injection control (UIC) program for carbon dioxide (CO₂) geologic sequestration (GS) wells. Final Rule, 75 FR 77230, December 10, 2010, https://www.federalregister.gov/documents/2010/12/10/2010-29954/federal-requirements-under-the-underground-injection-control-uic-program-for-carbon-dioxide-co2. 2010. Area-of-review modeling requirements in 40 CFR 146.84(a) and (c); accessed October 30, 2026.
- [27]Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291, 2023.arxiv.org/abs/2305.16291
- [28]Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, Ahmed H. Awadallah, Ryen W. White, Doug Burger, and Chi Wang. AutoGen: Enabling next-gen LLM applications via multi-agent conversations. In Conference on Language Modeling (COLM), 2024. arXiv:2308.08155.arxiv.org/abs/2308.08155
- [29]John Yang, Carlos E. Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, Karthik Narasimhan, and Ofir Press. SWE-agent: Agent-computer interfaces enable automated software engineering. In Advances in Neural Information Processing Systems (NeurIPS), 2024. arXiv:2405.15793.arxiv.org/abs/2405.15793
- [30]Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. ReAct: Synergizing reasoning and acting in language models. In International Conference on Learning Representations (ICLR), 2023. arXiv:2210.03629.arxiv.org/abs/2210.03629
