Engineering topics

Sedova, Ada

Publications and source records attributed to Sedova, Ada.

3 records

DOE OSTI2024

Impacts of floating-point non-associativity on reproducibility for HPC and deep learning applications

Run to run variability in parallel programs caused by floating-point non-associativity has been known to significantly affect reproducibility in iterative algorithms, due to accumulating errors. Non-reproducibility can critically affect the efficiency and effectiveness of correctness testing for stochastic programs. Recently, the sensitivity of deep learning training and inference pipelines to floating-point non-associativity has been found to sometimes be extreme. It can prevent certification for commercial applications, accurate assessment of robustness and sensitivity, and bug detection. New approaches in scientific computing applications have coupled deep learning models with high-performance computing, leading to an aggravation of debugging and testing challenges. Here we perform an investigation of the statistical properties of floating-point non-associativity within modern parallel programming models, and analyze performance and productivity impacts of replacing atomic operations with deterministic alternatives on GPUs. We examine the recently-added deterministic options in PyTorch within the context of GPU deployment for deep learning, uncovering and quantifying the impacts of input parameters triggering run to run variability and reporting on the reliability and completeness of the documentation. Finally, we evaluate the strategy of exploiting automatic determinism that could be provided by deterministic hardware, using the Groq LPUTM accelerator for inference portions of the deep learning pipeline. We demonstrate the benefits that a hardware-based strategy can provide within reproducibility and correctness efforts.

Shanmugavelu, Sanjif↗

DOE OSTI2024

Deep-Learning Interatomic Potential Connects Molecular Structural Ordering to the Macroscale Properties of Polyacrylonitrile

Polyacrylonitrile (PAN) is an important commercial polymer, bearing atactic stereochemistry resulting from nonselective radical polymerization. As such, an accurate, fundamental understanding of governing interactions among PAN molecular units is indispensable for advancing the design principles of final products at reduced processability costs. While ab initio molecular dynamics (AIMD) simulations can provide the necessary accuracy for treating key interactions in polar polymers, such as dipole–dipole interactions and hydrogen bonding, and analyzing their influence on the molecular orientation, their implementation is limited to small molecules only. Herein, we show that the neural network interatomic potentials (NNIPs) that are trained on the small-scale AIMD data (acquired for oligomers) can be efficiently employed to examine the structures and properties at large scales (polymers). NNIP provides critical insight into intra- and interchain hydrogen-bonding and dipolar correlations and accurately predicts the amorphous bulk PAN structure validated by modeling the experimental X-ray structure factor. Furthermore, the NNIP-predicted PAN properties, such as density and elastic modulus, are in good agreement with their experimental values. Overall, the trend in the elastic modulus is found to correlate strongly with the PAN structural orientations encoded in the Hermans orientation factor. In conclusion, this study enables the ability to predict the structure–property relations for PAN and analogues with sustainable ab initio accuracy across scales.

36 MATERIALS SCIENCE↗

DOE OSTI2024

High throughput, accurate gene annotation through AI and HPC-enabled structural analysis

With the advances in next generation sequencing technologies, the number of sequenced genomes is growing exponentially, resulting in a technology bottleneck for the translation of sequence information into usable hypotheses about the function of each gene. We have proposed leveraging our leadership high-performance computing (HPC) resources to help break this annotation bottleneck. Here we design an HPC-based framework to infer gene function from gene sequence by incorporating information about protein structure and interactions predicted by deep learning approaches. Accurate functional prediction and gene annotation using computational methods will facilitate breakthroughs in the genomic sciences essential to understanding and harnessing life processes in bacteria, fungi and plants. The development and applications of the state-of-the-art deep neural networks to protein structural modeling, interaction prediction, sequence comparison, and quality assessment of protein structural models will be made possible by leadership computational resources. These HPC-enabled bioinformatics and molecular modeling tools will provide powerful insights into molecular functions of genes.

59 BASIC BIOLOGICAL SCIENCES↗