Argonne National Laboratory is building scientific computing ecosystems that integrate AI across institutional, software, and community dimensions to enable data-driven discovery.

Trustworthy AI-enabled scientific discovery depends on software, people, institutions and community action.

Artificial intelligence is changing scientific computing rapidly. AI systems can now generate and analyze code, synthesize information, assist with hypothesis generation and carry out complex reasoning tasks within scientific workflows. Alongside advances in computing architectures and data-intensive research, these capabilities are reshaping how scientific work is conducted.

The challenge for the scientific community is not simply to make research faster, but to build scientific computing ecosystems that can harness these capabilities while preserving rigor, transparency, accountability and human judgment — enabling new forms of collaboration and discovery.

This was the focus of the 2026 Toward Next-Generation Ecosystems for Scientific Computing workshop, held April 14–16 in Chicago. Organized by Lois Curfman McInnes and collaborators as part of her 2024 DOE Office of Science Distinguished Scientist Fellowship, the workshop brought together researchers and practitioners from universities, national laboratories, industry and other organizations to consider how scientific computing ecosystems should evolve as AI becomes a more active participant in research. Building on the first workshop in 2025, this year's discussions moved from identifying challenges toward developing strategic priorities and actions for the community.

The resulting report identifies four interdependent strategic themes: software ecosystems for AI-enabled scientific discovery; trust, validation and traceability; human-AI teaming and paradigm shifts; and workforce, pedagogy and governance.

The report uses ecosystem in a broad sense, encompassing the technical, social and organizational structures that shape scientific computing — from hardware, software, data and workflows to people, institutions, incentives and community practices. The goal is to create an environment that can adapt as scientific needs and technologies change. This perspective is critical as AI evolves from being a tool used by researchers toward becoming an active participant in scientific workflows, raising pressing questions about validation, provenance, accountability and the role of human expertise.

Essential to effective software ecosystems are shared research assets: reusable software, datasets, metadata, workflows, benchmarks, validation examples and documentation that capture knowledge, context, assumptions and computational approaches in forms that can be reused across projects and communities. These assets help researchers understand, extend and validate complex workflows, particularly as AI, accelerator-based computing and emerging technologies become increasingly interconnected. The report emphasizes usability — helping researchers identify appropriate methods and their limitations for addressing scientific problems.

As AI-enabled workflows move from exploratory use to scientific analysis and decision-making, the report calls for trust-related capabilities — verification, validation, uncertainty quantification, provenance, traceability, transparency and auditability — to become core design requirements. Traditional approaches may not suffice; they must be extended with methods tailored to hybrid workflows that combine mechanistic models, numerical methods, AI systems, surrogate models and human judgment. Scientific software presents special challenges: it can run without obvious technical error and still generate scientifically misleading results, and conventional tests may not detect these situations, particularly when AI is rapidly generating or modifying code. Workshop participants stressed the need for new mathematically and statistically grounded approaches to evaluating AI-enabled scientific workflows.

The report distinguishes between taskwork — activities needed to accomplish a particular goal — and teamwork, which concerns how people and AI systems coordinate effectively toward that goal. As AI takes on more taskwork, effective teamwork becomes increasingly important. Human researchers remain essential for framing scientific questions, combining domain and computational knowledge, evaluating evidence, managing uncertainty and interpreting unexpected results. The appropriate balance between automation and human involvement depends on the task, with higher-risk or higher-consequence activities requiring greater transparency, validation and human judgment. The workshop emphasizes giving early-career researchers practical experience with human-AI teamwork through fellowships, internships, apprenticeships and cross-sector research projects.

Training for scientific computing must extend beyond programming to include critical thinking, evaluation of evidence, reasoning under uncertainty, verification, reproducibility, communication and responsible use of AI-enabled tools. The report's workforce roadmap emphasizes multiple pathways for students, practitioners and institutional leaders, recognizing that people enter scientific computing at different stages and in different roles. Rather than preparing people for a particular generation of tools, the goal is to develop competencies that remain valuable as technologies evolve.

The workshop's central message is that longstanding ecosystem challenges, intensified by AI, cannot be addressed by individual projects or organizations acting alone. Building effective scientific ecosystems requires coordination among researchers, universities, national laboratories, the private sector, funders and professional societies.

Based on the four themes, the report identifies eight areas for community action. Near-term efforts should strengthen shared research assets, trust infrastructure, human-AI teaming practices and incentives for scientific software stewardship. Pilot projects should explore new approaches for coordinating multiple AI agents and multiple institutions while investigating ways to help users make better-informed choices throughout research workflows. Longitudinal studies should examine how AI affects the workplace, professional roles and team behavior and evaluate whether AI-enabled ecosystems improve scientific and engineering practice.

AI can accelerate parts of scientific work, but speed alone is not progress. The relevant question is whether AI-enabled ecosystems support valid, interpretable and creative discovery. Meeting that standard requires more than technical capability: software, data, people, institutions, incentives and community practices must evolve together through feedback, adaptation and stewardship. As the report concludes, "Done well, these ecosystems can accelerate discovery and expand scientific inquiry while preserving the rigor, transparency, accountability and collaboration on which progress and public benefit depend."

The full report, "Toward Next-Generation Ecosystems for Scientific Computing: Harnessing Community, Software, and AI for Cross-Disciplinary Team Science" by L. C. McInnes et al., is available at https://doi.org/10.48550/arXiv.2608.26519. Those attending SC26 are invited to join the conversation Tuesday, November 17 at BOF: Ecosystems for Scientific Computing in the Age of AI.