Matthew Ellis

Riccardo Balin, Filippo Simini, Cooper Simpson et al.

Recent years have seen many successful applications of machine learning (ML) to facilitate fluid dynamic computations. As simulations grow, generating new training datasets for traditional offline learning creates I/O and storage bottlenecks. Additionally, performing inference at runtime requires non-trivial coupling of ML framework libraries with simulation codes. This work offers a solution to both limitations by simplifying this coupling and enabling in situ training and inference workflows on heterogeneous clusters. Leveraging SmartSim, the presented framework deploys a database to store data and ML models in memory, thus circumventing the file system. On the Polaris supercomputer, we demonstrate perfect scaling efficiency to the full machine size of the data transfer and inference costs thanks to a novel co-located deployment of the database. Moreover, we train an autoencoder in situ from a turbulent flow simulation, showing that the framework overhead is negligible relative to a solver time step and training epoch.

Michael Gao, Suresh Balu, William Knechtle et al.

Clinical documentation and data retrieval within Electronic Health Records (EHRs) contribute substantially to clinician workload and burnout. To address this, we developed Scout, an LLM-based EHR search and synthesis platform that enables clinicians to query EHR data using natural language. Each response includes citations linking each claim to the original data source, facilitating easy verification of generated content. We conducted a prospective randomized, evaluator-blinded crossover trial across seven clinical specialties (20 participants, 200 structured cases). Participants completed realistic clinical tasks using either Scout or the EHR alone, with outcomes including time to completion, NASA Task Load Index workload scores, and blinded expert adjudication of accuracy, completeness, and relevance. Scout reduced task completion time by 37.6% and significantly decreased perceived workload, with the largest reductions in mental demand, effort, and temporal demand. Non-inferiority analyses showed that tasks completed with Scout maintained accuracy, completeness, and relevance relative to tasks completed with the EHR-only. A concurrent pilot deployment across over 200 users and more than 20 specialties generated over 6,600 interactions in three months, revealing diverse clinical and administrative use cases. Automated evaluation using an LLM-as-judge framework identified errors at low rates. Subsequent manual review of a subset of outputs revealed that most claims flagged by the automated judge as errors were in fact supported by the patient chart, demonstrating the importance of human validation. These findings provide early trial-based evidence that LLM-powered EHR tools can meaningfully reduce clinical and administrative workloads while maintaining output quality.

Sam Partee, Matthew Ellis, Alessandro Rigazzi et al.

We demonstrate the first climate-scale, numerical ocean simulations improved through distributed, online inference of Deep Neural Networks (DNN) using SmartSim. SmartSim is a library dedicated to enabling online analysis and Machine Learning (ML) for traditional HPC simulations. In this paper, we detail the SmartSim architecture and provide benchmarks including online inference with a shared ML model on heterogeneous HPC systems. We demonstrate the capability of SmartSim by using it to run a 12-member ensemble of global-scale, high-resolution ocean simulations, each spanning 19 compute nodes, all communicating with the same ML architecture at each simulation timestep. In total, 970 billion inferences are collectively served by running the ensemble for a total of 120 simulated years. Finally, we show our solution is stable over the full duration of the model integrations, and that the inclusion of machine learning has minimal impact on the simulation runtimes.