PETSc AI Website#

Generative AI provides unprecedented opportunities to transform strategies for research and development (R&D) of scientific software, with the ultimate goal of accelerating scientific insight. PETSc—the Portable, Extensible Toolkit for Scientific Computation—provides foundational software for scalable numerical solvers, including related numerical optimization (TAO).

Vision for PETSc AI Agents#

Our goal is to create PETSc AI agents and an agent community capable of understanding and executing scientific computing tasks at the heart of numerical library research, providing the AI equivalent of human PETSc developers using LLM technology, while also continually advancing the PETSc knowledge base.

Key domains for PETSc AI agemts

This diagram illustrates key domains in which we envision LLM-powered agents to meaningfully assist PETSc R&D and user support while also augmenting the PETSc knowledge base for long-term community advances. We do not aim to replace human developers, but rather to extend their capabilities and improve productivity. These AI agents will operate within the collaborative infrastructure of PETSc—subject to the same review and git merge request processes as human contributors—to ensure that any changes to the official knowledge base are carefully vetted.

Recent Progress#

  • MCP tools for scientific computing with PETSc

    • Project info: https://mcp.petsc-ai.org

    • Initial emphasis: A common framework of basic MCP agents that provide building blocks for more advanced functionality

    • Reference: Improving Usability and Productivity of PETSc with Agent-Based Workflows, B. Smith, H. Zhang, J. Zhang, S. Balay, L. Chen, M. Keceli, L.C. McInnes, PEARC ‘26: Proceedings of the Practice and Experience in Advanced Research Computing 2026: Resilient Roots + Empowered Communities, Article 38, pages 1-5, https://doi.org/10.1145/3785462.3815882

  • PETScAgent-Bench

    • Framework for writing PETSc/TAO benchmarks and using them to evaluate output from LLM-based agents or environments

    • Project repo: petsc/petscagent-bench

    • News! Feb 2026: 3rd-place winner in the Code Generation track of the AgentX-AgentBeats competition

    • Reference: An Agentic Evaluation, Framework for AI-Generated Scientific Code in PETSc, Zhang H, Smith B, Balay S, Chen L, Keceli M, McInnes LC, Zhang J. ArXiv, March 2026, https://doi.org/10.48550/arXiv.2603.15976

  • AI assistants to enhance and exploit the PETSc knowledge base

    • Reference: AI Assistants to Enhance and Exploit the PETSc Knowledge Base, B. Smith, J. Zhang, H. Zhang, L.C. McInnes, M. Keceli, A. Vasan, S. Balay, T. Isaac, L. Chen, V. Vishwanath, Workshop on AI Assisted Software Development for HPC, in conjunction with the 54th International Conference on Parallel Processing, ICPP 2025, Sept 8-11, 2025, San Diego, CA, also available via https://doi.org/10.48550/arXiv.2506.20608

Next Steps#

Build on this foundation to create advanced AI capabilities as needed for next-generation science.

Note: Work on PETSc AI agents is being pursued by a subset of PETSc developers, including team members at Argonne National Laboratory.