Vision-based scientific foundation models hold significant promise for advancing scientific discovery and innovation. This potential stems from their ability to aggregate images from diverse sources—such as varying physical groundings or data acquisition systems—and to learn spatio-temporal correlations using transformer architectures. However, tokenizing and aggregating images can be compute-intensive, a challenge not fully addressed by current distributed methods. In this work, we introduce the Distributed Cross-Channel Hierarchical Aggregation (D-CHAG) approach designed for datasets with a large number of channels across image modalities. Our method is compatible with any model-parallel strategy and any type of vision transformer architecture, significantly improving computational efficiency. We evaluated D-CHAG on hyperspectral imaging and weather forecasting tasks. When integrated with tensor parallelism and model sharding, our approach achieved up to a 75% reduction in memory usage and more than doubled sustained throughput on up to 1,024 AMD GPUs on the Frontier Supercomputer.

This paper presents a novel deconvolution mechanism, called the Sparse Deconvolution, that generalizes the classical transpose convolution operation to sparse unstructured domains, enabling the fast and accurate generation and upsampling of point clouds and other irregular data. Specifically, the approach uses deconvolutional kernels, which each map an input feature vector and set of trainable scalar weights to the feature vectors of multiple child output elements. Unlike previous approaches, the Sparse Deconvolution does not require any voxelization or structured formulation of data, it is scalable to a large number of elements, and it is capable of utilizing local feature information. As a result, these capabilities allow for the practical generation of unstructured data in unsupervised settings. Preliminary experiments are performed here, where Sparse Deconvolution layers are used as a generator within an autoencoder trained on the 3D MNIST dataset.

Biologically-informed neural networks (BINNs), an extension of physics-informed neural networks [1], are introduced and used to discover the underlying dynamics of biological systems from sparse experimental data. In the present work, BINNs are trained in a supervised learning framework to approximate in vitro cell biology assay experiments while respecting a generalized form of the governing reaction-diffusion partial differential equation (PDE). By allowing the diffusion and reaction terms to be multilayer perceptrons (MLPs), the nonlinear forms of these terms can be learned while simultaneously converging to the solution of the governing PDE. Further, the trained MLPs are used to guide the selection of biologically interpretable mechanistic forms of the PDE terms which provides new insights into the biological and physical mechanisms that govern the dynamics of the observed system. The method is evaluated on sparse real-world data from wound healing assays with varying initial cell densities [2].