Norsk English

Agentic AI for scientific computing and simulation

We help research and industry connect AI agents to simulation software and develop workflows for modelling, analysis and decision support. Our work combines agent technology with expertise in numerical methods, simulator development and scientific software.

Contact persons

What we can help you with

Scientific and engineering work often involves moving between data, models, software and expert assessments. AI agents can help carry out these steps and connect them into an iterative workflow: interpreting a task, using tools, inspecting results and proposing what to do next.

We work with industrial partners and research organisations to:

  • Connect AI agents to simulators, data sources and specialist software.
  • Develop workflows for model construction, simulation, analysis and comparison of alternatives.
  • Build tailored interfaces and analysis tools around established scientific software.
  • Improve software interfaces, documentation and diagnostics so that agents can use tools effectively and recover from errors.
  • Evaluate agent performance and develop mechanisms for traceability, reproducibility and expert review.

Do you have a modelling task, a simulator you would like to connect to an agent, or a workflow that involves substantial manual work? Get in touch with our Applied Computational Science research team to discuss the possibilities. Together, we can identify where agents would be useful, develop and test a suitable solution, and establish how its results should be assessed.

Agents as collaborators in computational science

Our aim is to make agents useful collaborators in demanding technical work: something an expert can discuss a problem with, delegate tasks to and receive informed suggestions from. This requires agents that can act on feedback, revise their approach and recognise when they need clarification.

In an agent-assisted modelling workflow, the user describes a problem and its objectives. The agent can retrieve documentation, construct a model, write and execute code, inspect solver diagnostics and analyse outputs. The user and agent then refine assumptions, compare alternatives and decide which questions to investigate next.

We develop and investigate these capabilities through tools such as JutulAgent, which connects language models to scientific simulators built on the Jutul framework. This provides a practical environment for studying how people and agents can work together, and what an agent can reliably take responsibility for.

Connecting models, data and specialist tools

Many technical questions span several applications. A geological interpretation may provide the basis for a flow model; environmental data may define inputs to a flood simulation; simulation results may inform an optimisation problem. Moving between these steps often requires manual conversion, scripting and coordination between specialists.

Agents can help connect such tools, prepare inputs and carry information between stages. The scientific challenge is to preserve the meaning of the data and assumptions throughout the workflow. Units, boundary conditions, model scope and uncertainty must remain explicit as information passes from one application to another.

We develop these connections using software interfaces and open standards, including the Model Context Protocol (MCP). The scope of automation and the checks required are determined by the application and the decisions the workflow is intended to support.

Building software around simulation

Agentic coding also makes it possible to develop interfaces and supporting tools for a particular task. Examples include interactive visualisation, data preparation, scenario comparison and automated reporting around an existing simulator.

Our approach builds on established numerical software while using agents to develop and adapt the surrounding tools. This creates opportunities to make specialist methods accessible in new settings and to explore ideas that previously required substantial implementation effort.

The resulting software still needs appropriate testing and review. Scientific expertise is essential both for choosing useful functionality and for checking that the implementation preserves the intended behaviour of the underlying models.

Why simulator expertise matters

An agent needs feedback it can act on. Consistency checks, well-defined interfaces and informative solver diagnostics help it detect invalid inputs, investigate failed calculations and revise its approach. Documentation and software design therefore directly affect how effectively an agent can work.

SINTEF combines expertise in numerical methods and simulator development with research on agent technology. We work on both sides of the interaction: how agents use scientific software, and how that software can provide better feedback to agents. Differentiable simulators can also supply sensitivities that support parameter estimation, optimisation and systematic exploration of a model.

A completed calculation does not establish that the model represents the intended problem or that its predictions are adequate for a decision. Assessing physical assumptions, numerical accuracy and the relevance of the input data remains a central part of the workflow.

Traceability and expert judgement

A short instruction leaves many choices unstated. An agent may select parameter values, boundary conditions or modelling assumptions that the user did not explicitly specify. These choices must be visible if the user is to assess whether the agent has solved the right problem.

We develop workflows that record tool calls, generated code and outputs, and make assumptions and modelling choices available for review. Reproducing a calculation also requires retaining the relevant input data, software versions and execution settings.

The division of responsibility between people and agents is part of the design. Routine steps may be delegated within an agreed scope, while consequential choices or unresolved ambiguities require expert input. The appropriate balance depends on the task, the evidence available and the consequences of an incorrect result.

Towards more independent scientific work

Our longer-term research ambition is for agents to contribute across the scientific cycle: formulating hypotheses, designing computational experiments, interpreting results and revising models. Progress depends on their ability to assess their own work, maintain a coherent direction across many tasks and recognise when to ask for help.

We investigate these questions through concrete applications and systematic evaluation. As agent technology evolves, our work builds on lasting foundations: numerical methods, domain expertise, robust simulation software and the ability to inspect and verify results.

We welcome collaboration on agent-assisted modelling, connections to existing scientific software, and research into more capable and reliable computational workflows.

Read more