AI-Accelerated Science
Science begins with curiosity, the questions we ask about the world we are part of. For most of my career, every answer to a new question waited weeks or more while I executed each step. Today, AI helps me carry analysis from idea to result in hours and my time returns to the creative, detailed, and exploratory thinking that pushes the edge of what we can do as scientists. When execution no longer stands between a question and its answer, what remains at the center of science is the scientist's judgment: which questions matter, how to pursue them, and what their answers mean.
This judgment is what turns AI-accelerated work into science. I architect mine into the software and structure that AI models and I work within. I hold and guide the what, why, and how: I decide why, we work out how together, and the model composes what. The boundary between us moves as models improve. SolarWindPy is built this way. Its design separates the physics from the code, so I can give each my full attention when it needs it. I am now architecting it so an AI model can maintain it with less direct input. During an AI-assisted release sprint in 2025, I committed changes to SolarWindPy nearly six times as often per working day as in the decade before. Each release is tested, documented, and permanently archived, so anyone can use it, cite the exact version behind a result, and learn the scientific reasoning built into it. Paying that up-front cost once means each new analysis starts and reaches further ahead.
AI-accelerated science is valuable when we can trust its results. I work with AI models as collaborators: a model checks my work and challenges my reasoning, as I check and challenge its, so that what we deliver together is stronger than what either of us would produce alone. Statistical tests and error estimates that once took me days or weeks now take a fraction of that time, and a model can test an idea while I continue advancing the physics. This rigor tells me which research directions to pursue and which to set aside with confidence. I apply the same judgment to my tools, testing each one by comparing how models work with and without it and calibrating it against my own expertise, which lets me adapt it as models change. Held to this standard of accountability, acceleration strengthens the trust our community can place in our published results, code, and data.
Scientific judgment lives in people, and it grows through practice, in the teams I build and the students I mentor. With students, I build the scaffolding that gives them the freedom to explore, question, and develop their own scientific intuition. I push them to challenge my reasoning, because learning to question is how judgment grows. The same practice keeps refining my own judgment about what can move from the scientist to the model or its tools, and what must stay with us. As AI takes on more of the execution, the ability to ask and to challenge becomes the heart of a scientist's training, and it keeps judgment growing in the people who will carry science forward.