Discovery Through AI: A New Era of Scientific Computing

Achieving scientific breakthroughs such as full-scale cellular modeling, whole-earth climate simulation, accelerated materials discovery, and digital twins will require orders-of-magnitude increases in computing capability and researcher productivity. High-performance computing and artificial intelligence are converging to make that possible: with the  term   "AI for Science". AMD is enabling AI for Science by fusing leadership-class compute, open software, and domain partnerships to accelerate discovery while preserving the rigor of high-fidelity modeling and simulation. 

At Oak Ridge National Laboratory (ORNL), two newly announced systems exemplify this strategy. The Lux AI supercomputer, powered by AMD and deploying in early 2026, will be the first US AI factory for science. Lux is designed to expand US Department of Energy (DOE) AI leadership and accelerate breakthroughs across energy, materials, medicine, and advanced manufacturing. The Lux system will be based on AMD EPYC CPUs, codenamed Turin, and AMD Instinct MI355X GPUs, with the AMD Pensando Pollara Network Cards. In addition to this unique hardware platform, a key differentiator is its AI-factory model delivered on-premises with cloud services: Lux will host AI capabilities using open-source orchestration and microservices. The AMD AI Enterprise Suite underpins this, enabling elastic, multi-tenant AI workflows, and supporting heterogeneous clusters, so researchers can integrate diverse resources without re-architecting their software.

Discovery, the next-generation supercomputer at ORNL, deepens collaboration between the DOE, ORNL, HPE, and AMD to advance  US AI and scientific research at massive scale. Discovery will be  powered by next-gen AMD EPYC CPUs, codenamed “Venice,” and AMD Instinct MI430X Series GPUs:  engineered for sovereign AI and scientific computing. Together, these systems will help the  US train, simulate, and deploy AI models on domestically built infrastructure, safeguarding data and competitiveness while accelerating AI-enabled science. 

AMD believes AI will significantly augment, but not replace, traditional modeling and simulation (ModSim). AI for Science delivering orders-of-magnitude gains in capability and productivity through increased automation, federation, and AI acceleration. For example, AMD and Lawrence Livermore National Laboratory (LLNL) have demonstrated AI-driven biology at unprecedented scale using ElMerFold for protein structure prediction, running on the entire system of El Capitan. Another exciting AI for Science application was detailed at ORBIT-2: AMD and ORNL Advancing Earth System Intelligence at Exascale. These efforts show how advanced hardware, open software, and co-design with domain experts translate into practical, scientific impact.

Complimenting Scientific Simulation with AI

AMD envisions AI will complement traditional modeling and simulation through several motifs:

• AI-augmented simulations: Emerging workflows interweave FP64-heavy ModSim with low-precision inference for embedded surrogate models and real-time orchestration. This demands nodes capable of both high-throughput FP64 and efficient AI inference/training, as well as cluster-wide orchestration models that can launch and steer simulations at scale.

• Acceleration via surrogates and mixed precision: AI surrogates can replace or coarsen expensive kernels, while mixed/reduced-precision datatypes accelerate throughput. These surrogates speed sweeps, sensitivity studies, and uncertainty propagation, freeing FP64 computations for the places that require the highest-fidelity solutions.

• Informing simulations with digital twins and inverse design: Digital twins tightly couple live data to models, enabling predictive control and rapid what-if exploration. Inverse design uses generative models and optimization models to navigate vast parameter spaces, accelerating the discovery of materials, devices, and processes.

• Autonomous computational operations: Agentic workflows manage end-to-end loops across data acquisition, simulation, inference, and validation, closing the gap between hypothesis generation and verification.

Crucially, AMD continues to invest in native, IEEE-compliant FP64 arithmetic with high performance. Many high-consequence simulations require full double precision due to wide dynamic ranges, ill-conditioned systems, and chaotic dynamics. High-fidelity FP64 remains essential for generating training data, validating AI predictions, and anchoring reduced-order models in first principles.

Digital twins represent a particularly promising modality. By linking experimental facilities and streaming sensors to HPC/AI surrogates, researchers can enact real-time calibration and decision-making. These hybrid workflows depend on scalable I/O, latency-aware scheduling, and robust model serving, which includes capabilities delivered by the AMD AI Enterprise Stack with open-source orchestration and microservices for portable, reproducible deployments across on-prem, cloud, and hybrid environments.

Inverse design expands beyond classic optimization to strengthen verification, validation, and uncertainty quantification (V&V/UQ). Outer-loop AI automation can massively parallelize sampling, intelligently explore parameter spaces, and systematically probe edge cases. Generative models can prioritize simulations, guide optimal experimental design, and adapt search strategies. These are precisely the kind of facility-scale automation Lux’s AI factory model and Discovery’s extreme compute are built to accelerate.

The Future of High Performance Computing

Scientific computing is converging on three pillars: Modeling & Simulation for resolved physics at scale, AI for surrogate-driven acceleration and automation, and Quantum for specialized speedups where classical approaches are intractable. Their integration on open, heterogeneous platforms creates a resilient pathway from simulation to experiment to deployment. By combining sovereign-scale systems (Discovery), cloud-native AI factories (Lux), and an open, portable software stack (ROCm) that embraces heterogeneity, AMD is enabling a new era of AI-enabled scientific discovery.