AI is transforming scientific research by acting as a co-scientist that collaborates with human experts to accelerate discovery. These AI co-scientists are multi-agent systems that generate, test, and refine scientific hypotheses, design experiments, analyze data, and provide insights through advanced reasoning and interdisciplinary knowledge. Powered by large language models (LLMs) and trained on vast scientific datasets, they can identify hidden patterns, connect unrelated fields, and propose new research directions—enabling faster, more reproducible science. NVIDIA is at the forefront of this revolution, supporting the development of AI co-scientists at Los Alamos National Laboratory (LANL) to tackle two of the most complex challenges in science: inertial confinement fusion (ICF) and targeted cancer therapy. In the field of fusion energy, LANL and NVIDIA are building an AI co-scientist to help solve the intricate physics of ICF. Fusion, the process that powers stars, involves compressing and heating a fuel pellet with lasers to trigger nuclear fusion. However, the process is highly complex, involving multiple physical phenomena across vastly different scales, making accurate predictions difficult. Experimental results often differ from simulations due to subtle changes in initial conditions or design parameters. To address this, LANL is adapting the NVIDIA NeMo framework and the Llama Nemotron Super 1.5 model. The process includes domain-adapted pretraining using open-access scientific documents from arXiv, CORE, and OSTI.gov, followed by supervised fine-tuning and reasoning trace training. The resulting model is evaluated using expert-generated benchmarks to ensure scientific accuracy. The AI co-scientist then generates hypotheses, runs simulations, and iteratively refines designs based on experimental feedback from facilities like the National Ignition Facility and OMEGA. This accelerates progress toward reliable fusion energy and deeper understanding of extreme matter. In cancer research, the team is developing an AI co-scientist for targeted alpha therapy (TAT), a promising treatment that delivers radioactive alpha particles directly to tumors. The success of TAT depends on chelator molecules that bind and transport radioactive atoms precisely to cancer cells. However, designing effective chelators is difficult because the metals used have large atomic radii, and few stable binding molecules are known. To overcome this, LANL is using a hybrid AI workflow powered by NVIDIA. The process begins with Llama Nemotron Super 1.5 generating scientific hypotheses about ideal chelator properties. These hypotheses are then used to guide GenMol, a generative AI model that designs new molecular candidates. The proposed molecules are combined with radioactive atoms using Architector, and their 3D structures are simulated on LANL’s Venado supercomputer using high-performance quantum chemistry methods. The simulation results are fed back into the AI to validate or refine the hypothesis, closing the loop for rapid iteration. This approach has already led to the discovery of chelators with improved binding energy for actinium, a key element in TAT. The method not only speeds up molecular discovery but also reveals the chemical traits that make certain molecules more effective. The technology has broader applications in treating poisoning, metal purification, and other chemical processes. Mark Chadwick, Associate Laboratory Director for Simulation, Computing, and Theory at LANL, emphasized the transformative potential: “With NVIDIA, we are pioneering AI co-scientists that combine domain expertise with cutting-edge AI to solve humanity’s grand challenges.” This work was supported by the Perlmutter supercomputer at NERSC, a U.S. Department of Energy user facility. Scientists interested in building their own AI co-scientists can explore NVIDIA NeMo and Nemotron to get started. For more details, join LANL and NVIDIA at SC25, where they will present on AI in fusion and molecular discovery.