A Conversational User Interface for Autonomous Reservoir Simulation Deck Generation and Execution

Abstract

Abstract Objectives/Scope This paper presents CLARISSA (Conversational Language Agent for Reservoir Integrated Simulation System Analysis), a conversational agent that autonomously generates, executes, and analyzes reservoir simulation models from natural language input. Engineers describe what they want; CLARISSA delivers working models. To enable standardized comparison of conversational deck-generation systems, we introduce RIGOR (Reservoir Input Generation Output Review), a benchmark for evaluating CLARISSA and future similar efforts. Methods, Procedures, Process CLARISSA employs large language models integrated with OPM Flow through a multi-stage translation pipeline that converts natural language problem descriptions into PetroScript, a typed, simulator-independent authoring language whose compiler currently emits Eclipse-format deck syntax. The system supports multiple input modalities: voice, text chat, web interfaces, and REST APIs. Physics-aware validation flags inconsistencies during conversational elicitation, while analog databases suggest reasonable defaults for incomplete specifications—all assumptions are explicitly documented. Generated decks execute autonomously on OPM Flow; outputs can be exported for commercial platforms when third-party validation is required. RIGOR employs execution-based evaluation using simulator diagnostics including convergence behavior and material balance closure. Results, Observations, Conclusions CLARISSA successfully generated decks for SPE comparative solution projects including SPE1 (Odeh, 1981), SPE5 (Killough and Kossack, 1987), and SPE9 (Killough, 1995) from text and tabular input specifications. Generated decks executed without error and matched published benchmark results. Beyond validation cases, CLARISSA demonstrated autonomous model reformulation: given only a compositional table, the system converted a black-oil model to a fully compositional equation-of-state formulation without additional user intervention. This capability enables rapid evaluation of tertiary recovery options from existing assets. RIGOR defines complexity tiers from foundational (linear displacement corefloods) through intermediate (multi-well pattern floods) to advanced scenarios requiring mid-conversation compositional pivots. Execution-based scoring derives from simulator diagnostics: convergence behavior, material balance closure, and physical constraint satisfaction. The binding constraint on simulation adoption has never been solver performance—it is human cognitive load and workflow friction. CLARISSA eliminates both commercial licensing barriers and the specialist training traditionally required, enabling practicing reservoir engineers to directly leverage simulation capabilities previously accessible only to dedicated specialists. Novel/Additive Information Unlike prior work focused on querying existing models (SPE-221987) or retrieving reservoir data through pre-built tool functions (SPE-229646), CLARISSA is the first system capable of generating—not merely querying or modifying—complete simulation decks from natural language. RIGOR provides the first standardized benchmark for evaluating conversational deck-generation systems, enabling reproducible comparison of future approaches analogous to established benchmarks in software engineering code generation. PetroScript, to be released alongside RIGOR, is to our knowledge the first typed authoring language for reservoir simulation decks designed to be simulator-independent, replacing positional keyword syntax with named, validated parameters.