Ian Cloët, Deputy Division Director and Group Leader at Argonne National Laboratory, explains how Argonne’s Physics division is using AI safely and responsibly to enhance its work
Embedded in many aspects of modern life, artificial intelligence (AI) is transforming the way we work, learn, and innovate. This is particularly true for the world of physics, with AI capabilities offering new ways to investigate the fundamental laws that govern the Universe. From analysing vast datasets generated by particle accelerators and astronomical observations to simulating complex physical systems and predicting the behaviour of matter, AI is enabling researchers to tackle problems that were once computationally impractical or too complex to solve using traditional methods. However, the growing role of AI also raises important questions about the reliability, interpretability, and limitations of these techniques.
At Argonne National Laboratory, researchers are harnessing the power of AI to help advance their research across a broad range of disciplines. From engineering next-generation materials to speeding up breakthroughs in medical treatments, Argonne scientists are using AI to unlock new frontiers of discovery while enhancing resilience and competitiveness across critical sectors. To discover how AI is benefiting physics work at Argonne, The Innovation Platform spoke with Ian Cloët, Deputy Director of the Physics division and Group Leader of the division’s Theory group at Argonne National Laboratory.
How has AI changed the way physicists work in recent years?
Adoption of AI tools across the physics community has been widespread. However, its impact has been varied. Many researchers use generative AI tools to find answers to questions quicker than traditional literature or Google searches, whilst others rely heavily on these tools for coding and similar tasks, and some use AI to develop and orchestrate full end-to-end simulations of physics analysis and workflows.
AI has become an integral part of physics research and the extent to which it is being used is growing rapidly. As these tools become more powerful, and scientists develop the skills to develop and utilise agents, I expect these tools will soon revolutionise scientific research.
How is AI helping to accelerate research at Argonne National Laboratory?
Argonne is a very rich research environment, with many hundreds of scientists, major experimental facilities, and world-leading advanced computing. This combination allows AI tools and their development to be put to use in many different scenarios.
In nuclear physics, this includes improving the transport of ion beams at the Argonne Tandem Linac Accelerator System (ATLAS) – a U.S. Department of Energy user facility where hundreds of researchers from around the nation and the world conduct experiments each year. AI is also helping scientists analyse experimental data from ATLAS and other facilities worldwide, including Jefferson Lab and CERN. Research in nuclear theory is also a heavy user of AI, applying it to Lattice QCD calculations and the quantum many-body problem through collaborations such as STREAMLINE.
Currently, some of the most intensive use cases for AI are around automation, which allows for early detection of any issues with equipment, and the execution of standard but complicated workflows which dramatically increase productivity. However, increasingly, AI is being used to directly address frontier research questions, such as the calculation and extraction of 3D images of the quarks and gluons inside protons and nuclei. Across these applications, the goal is to enhance both efficiency and precision while enabling scientists to tackle increasingly complex research problems.
Can you share some specific examples of how AI is accelerating your work?
Here are three specific examples of how AI is aiding physics work at Argonne:
- Imaging the three-dimensional quark and gluon structure of the proton. The SciDAC-5 project ‘QuantOm’ (Quantum Chromodynamics Nuclear Tomography) is developing AI methods to take ‘pictures’ of the quarks and gluons that are the elementary building blocks of most visible matter in the Universe – that is, the proton and nuclei. AI tools are being used to check theory implementations and automate the extraction of these images from data, with the ultimate goal of real-time analysis at facilities such as Jefferson Lab and the forthcoming Electron-Ion Collider – a massive new particle accelerator facility under construction at Brookhaven National Laboratory. This data-driven approach helps address foundational questions such as: How does the strong force generate the vast majority of visible mass in the Universe? And why are quarks and gluons confined inside protons and other hadrons and can thereby never be detected in isolation?
- Understanding atomic nuclei and the quantum many-body problem. Here, neural networks are used to represent quantum wave functions, which provide mathematical descriptions of how protons and neutrons interact inside nuclei. This work aims to extend the reach of accurate nuclear many-body calculations. A recent application calculated the response of light nuclei to photons and tested the method against established numerical techniques as part of the STREAMLINE collaboration.
- Improving radioactive ion beam delivery at ATLAS. Argonne researchers have demonstrated an AI-assisted method that learns from measurements to select promising accelerator settings. This work shows advantages over conventional tuning in practical radioactive ion beam transport scenarios. More efficient tuning can make additional experimental time available for studying rare nuclei. arxiv.org
In addition, I use AI to more rapidly understand new concepts and research tools, to automate processes, validate results, write and check code, and to explore research ideas. This process significantly shortens the cycle from a research idea to an implemented and tested calculation.
Do you have to put particular provisions in place to ensure that AI is used responsibly, accurately, and safely?
Absolutely. In the context of research, safety is still crucial, particularly in regard to data protections. Unpublished work or data should not be shared with AI, where protections appropriate to the data are not built in.
AI is changing a lot of things and will completely revolutionise how research is performed. For example, most graduate students will be able to produce far more results during their graduate studies. However, this increase in productivity does not automatically translate into more knowledge and a deeper understanding.
With the advent of AI, it is critical that key aspects of the scientific culture are maintained, such as scientific integrity and not claiming ownership of work that you do not fully understand, cannot critically assess or defend, and did not play an active role in the development of. Maintaining these scientific standards for established scientists and instilling them in early career researchers will be critical for the responsible usage of AI.
What standards of validation are required before AI-generated results become publishable science?
AI is a tool, albeit a very powerful one, and does not replace the scientist as the person who takes ownership of any published result. Therefore, the standards of validation are consistent and even more stringent than those pre-AI, with additional checks appropriate to AI-specific failure modes.
Many uses of AI focus on automation and productivity improvements using standard workflows, so these results can be validated using previous methods.
In fact, AI provides new tools to check the validity of the workflows themselves, and using multiple AI systems can provide additional levels of validation that were not available or feasible before.
There are many methods physicists use to validate their own results, and these apply equally to AI-facilitated/generated results. These include:
- Ensuring the results satisfy the underlying laws of physics, such as relevant conservation laws, symmetries, known limits, and agreement with established calculations where available;
- Checking closure tests, where a complete end-to-end analysis is performed on simulated data and then ensuring a known ground truth is recovered with expected uncertainties;
- Making sure the results are robust against model assumptions and training choices, and understanding any edge cases; and
- Documenting the workflow and making sure the results are reproducible.
AI enables these checks to be more comprehensive as they reduce the effort for some validation steps. Adopting AI in research should lead to increased robustness of research results. In addition, AI makes it easier for scientists to validate aspects of the research results from their colleagues, so cross-checks across research communities are becoming much more common.
Please note, this article will also appear in the 28th edition of our quarterly publication.
