Reservoir Simulation Is the Proving Ground. Scientific AI Is the Opportunity.
CLARISSA · 27 Sep 2026
CLARISSA started with reservoir simulation for a reason.
Not because reservoir engineering is the only place we think agentic AI will matter.
Almost the opposite.
Reservoir simulation is an unusually demanding environment in which to find out whether scientific AI actually works.
A reservoir model combines specialized domain knowledge, structured data, numerical physics, formal input languages, multiple software tools, incomplete information, engineering judgment, and outputs that can be objectively tested.
That makes reservoir simulation a useful proving ground.
Reservoir simulation is where we are proving the architecture. The larger opportunity is scientific AI.
A harder test than answering questions
Much of today's AI interaction follows the same pattern:
A person asks a question.
A model generates an answer.
That pattern is enormously useful, but many scientific and engineering problems require something different.
An engineer does not merely need an explanation of a reservoir model. The engineer may need the system to modify that model, run it, evaluate the results, discover that something is wrong, repair the problem, rerun the case, document the assumptions, and preserve enough provenance that another engineer can understand what happened.
That requires an AI system capable of acting across the boundary between human intent and scientific software.
CLARISSA was designed around exactly that boundary.
The reservoir engineer states intent in professional language. CLARISSA converts that intent into an executable model and supports the result with simulation and deterministic validation rather than asking the engineer to trust an AI-generated assertion.
This leads to a pattern that is potentially much larger than one simulator:
Professional intent → structured scientific workflow → trusted computational tool → deterministic validation → human acceptance
Reservoir simulation gives us ground truth
Scientific AI has a difficult evaluation problem.
If an AI writes a polished explanation, another AI can grade the prose.
But did the scientific work actually succeed?
Reservoir simulation gives us something much stronger than stylistic judgment: a numerical simulator.
That is one reason RIGOR was created.
Instead of asking whether a generated reservoir model appears convincing, RIGOR can execute it and compare its behavior with reference physics. Its scoring framework emphasizes compilation, execution, convergence, numerical agreement, critical engineering KPIs, and process evidence.
That concept — scientific agents evaluated by the systems they are supposed to operate — is important well beyond reservoir engineering.
A chemistry agent should eventually be evaluated against chemistry.
A structural-engineering agent should be evaluated against structural models.
A geospatial agent should be evaluated against measurable terrain, mapping, and physical constraints.
Scientific AI needs more than eloquence.
It needs ground truth.
We are already moving beyond the reservoir simulator
The next expansion is not hypothetical.
CLARISSA currently focuses on reservoir engineering and expects a completed geomodel as part of its input. Initial testing is now extending the system into petrophysical analysis, geomodel analysis, and reservoir engineering as a connected workflow.
That progression is important.
Today, much of the subsurface workflow is divided across specialists and software packages:
Petrophysical interpretation produces rock and fluid properties.
Geomodeling organizes geological interpretation into a three-dimensional representation of the subsurface.
Reservoir simulation takes that representation and predicts dynamic behavior.
Each handoff creates another translation boundary.
If agentic AI can reliably operate across those boundaries while retaining provenance, deterministic checks, and human review, the opportunity becomes much larger than automating reservoir-deck syntax.
It begins to look like an agentic subsurface modeling system.
Why the architecture can travel
The most important parts of CLARISSA are not specific reservoir keywords.
They are architectural choices.
The system deliberately uses language models where semantic interpretation is required while assigning mechanical transformations to deterministic code. The simulator itself remains the syntax authority. Independent translation approaches can be compared. Physical and numerical checks occur before the result is accepted. Human engineering judgment remains the final gate.
PetroScript illustrates another part of the idea.
Instead of allowing an AI model to directly manipulate a fragile positional input syntax without constraints, PetroScript creates a typed semantic authoring surface and compiles the result deterministically into simulator input.
In reservoir engineering, that constrained layer is PetroScript.
In another scientific domain, the exact language or schema may be completely different.
The principle remains useful:
Do not ask an AI to improvise where software can enforce the rules.
Where could this eventually lead?
Our demonstrated capability today is reservoir engineering.
Our next active expansion is into petrophysics and geomodel analysis.
Beyond that, we see a broader class of problems in which the same approach may be useful: geothermal modeling, carbon storage, critical-mineral and mining workflows, groundwater, geoscience, geospatial engineering, and other domains built around complex scientific software.
Those are roadmap opportunities, not claims of capabilities already delivered.
That distinction matters.
We do not believe the right way to build scientific AI is to announce that one agent can suddenly perform every scientific discipline.
The more credible path is to prove the architecture in one difficult domain, expand into adjacent disciplines, measure what transfers, discover what does not, and repeat.
Reservoir simulation is that first difficult domain.
Scientific AI should inherit engineering's standards
Engineering software has developed over decades around concepts that AI systems are now rediscovering: unit tests, verification, validation, numerical tolerances, provenance, version control, input constraints, reproducibility, and human sign-off.
Those practices should not disappear simply because the interface becomes conversational.
If anything, autonomous scientific systems require them more.
CLARISSA's current architecture therefore separates conversational reasoning from the mechanisms responsible for accepting a model. A generated result must ultimately terminate in something independently inspectable — a simulator parser, a conservation identity, an analytical solution, a documented source, or another artifact an engineer can verify.
That is the direction we believe scientific AI has to go.
Not AI that merely talks about technical work.
AI that can participate in the work while remaining accountable to the tools, mathematics, physics, and people that determine whether the work is correct.
The bigger idea
Reservoir engineering is our starting point because it forces the difficult questions early.
Can the agent translate real engineering intent?
Can it operate specialized scientific software?
Can it deal with incomplete information?
Can it repair something it did not originally build?
Can another system independently measure whether it succeeded?
Can an engineer audit what happened afterward?
Those are not uniquely petroleum questions.
They are scientific-AI questions.
And that is why we think the opportunity is much larger than reservoir simulation.
Reservoir simulation is the proving ground. CLARISSA is the platform.
AI should not just talk about science. It should work with the tools that do science.