Loading…
Pipelining text and images
DOE Genesis Mission · Phase II
Multi-agent AI Expert for Subsurface Reasoning and Optimization
MAESTRO aims to coordinate AI agents, observations, physics-based simulation, and uncertainty to help researchers understand geothermal reservoirs and evaluate safer development choices.
The project
Enhanced geothermal systems (EGS) seek to recover heat from deep rock by creating and managing pathways for fluid circulation. Understanding those pathways, and how they change, is central to reservoir development.
MAESTRO—Multi-agent AI Expert for Subsurface Reasoning and Optimization—will bring observations, physical simulations, and scientific AI into a shared research workflow. Its goals include better reservoir characterization, informed stimulation and operations, and more explicit treatment of uncertainty and induced seismic risk.
Specialist models will infer fracture geometry, connectivity, and evolution alongside in-situ stress and reservoir dynamics. Coupled thermal, hydraulic, mechanical, and chemical (THMC) processes will provide the physical context.
A multimodal large language model (LLM) will coordinate physics-specialist agents. Physics-informed neural networks (PINNs), neural operators, diffusion sampling, and Bayesian data assimilation will help connect measurements with subsurface models.
The research will explore how knowledge from data-rich reservoirs can inform sites with fewer observations. Few-shot transfer and explicit uncertainty will help researchers examine which relationships carry across geological settings.
Continuous microseismic observations will support fracture tracking and research on induced seismic hazard. The aim is to inform stimulation and reservoir operations with clearer evidence about evolving conditions and containment.
Scientific workflow
Rather than treating a model answer as the end of an analysis, the proposed workflow will use uncertainty to guide the next observation, simulation, or check.
The planned evidence base combines continuous observations from 20 instrumented EGS sites with geoscientific literature and multi-fidelity simulation results.
GEOS, PFLOTRAN, EGS-Sandbox, and SW4 simulations will connect reservoir processes across scales, drawing on DOE exascale computing capabilities.
The orchestrator and specialist agents will compare observations and model predictions, using uncertainty to identify questions that need additional evidence.
Active learning will request targeted simulation ensembles and learn from their results. Conservation laws and code checks will help expose physically implausible answers.
Planned research interfaces
During Phase II, the project aims to develop interfaces that connect subsurface questions with data, model results, and their uncertainty.
Conversational interface
A conversational operator interface will help users investigate reservoir questions and review the observations, simulations, and physical reasoning behind proposed answers.
Geospatial workspace
A geospatial interface will support geothermal screening and data review, helping users explore subsurface evidence and uncertainty in its geographic context.
The longer-term aim is to reduce uncertainty, upfront capital risk, and development time while supporting better-informed geothermal planning.
The partnership
UC Irvine leads a collaboration spanning five University of California campuses, four national laboratories, and three industry partners. EnviTrace participates as an industry partner.
Connect with EnviTrace
Contact our team about MAESTRO, scientific collaboration, or related AI and subsurface research.