AI Compresses Scientific Data Without Forgetting

Massive scientific simulations, such as those modeling turbulent fluids or neural activity, generate data too large to store entirely, yet researchers need accurate analysis on the fly. A new AI approach tackles this by compressing data as it streams from simulations, ensuring high fidelity without requiring extensive memory, which could accelerate discoveries in fields like climate science and aerospace engineering.

The researchers developed a method that trains implicit neural representations (INRs)—AI models that learn continuous functions from data points—to compress scientific data in real-time, a process they term 'situ training'. This approach prevents 'catastrophic forgetting', where AI loses knowledge of earlier data, by using a combination of full and sketched data samples stored in limited buffers. As shown in Figure 3, the system employs buffers to retain recent snapshots and sketched versions of past data, allowing the AI to maintain accuracy over long simulation horizons.

The methodology relies on a novel regularization technique inspired by Johnson-Lindenstrauss transforms, which project high-dimensional data into lower dimensions while preserving essential relationships. Specifically, the team used fast Johnson-Lindenstrauss transforms (FJLT) and simple subsampling to create compressed sketches of data snapshots. These sketches, stored alongside full samples in buffers, serve as proxies for past data during training, enabling the AI to learn continuously without accessing the entire dataset. The training protocol, detailed in Algorithm 1, involves optimizing the INR model with mini-batches drawn from these buffers, balancing current and historical data to minimize errors.

Results from experiments on diverse datasets—Ignition (a 2D combustion simulation), Neuron (a diffusion process on an unstructured mesh), and Channel (a turbulent flow database)—demonstrate strong performance. For instance, in the Ignition dataset, the situ training with FJLT sketching achieved a peak signal-to-noise ratio (PSNR) of 37.4 dB and a relative Frobenius error (RFE) of 0.75%, matching the accuracy of offline methods that use the full dataset. Figure 5 and Figure 6 show visual comparisons where reconstructions closely mirror originals, even with compression rates up to 1882×. The study reports that sketching-based methods, particularly FJLT, outperformed subsampling in non-Cartesian geometries like the Neuron dataset, reducing errors and avoiding catastrophic failures.

This innovation matters because it enables real-time data compression in resource-constrained environments, such as running simulations on limited hardware or in distributed systems. For example, engineers could use this to monitor fluid dynamics in jet engines or climate models without delays, improving efficiency and decision-making. The mesh-agnostic nature of the approach means it applies to various data types, from uniform grids to complex unstructured meshes, broadening its utility in scientific and industrial applications.

Limitations include the dependence on sketch size and buffer capacity, which affect performance; smaller sketches may not fully capture data nuances, leading to higher errors in certain scenarios. The paper notes that the theoretical guarantees assume the data lies on a low-dimensional manifold, but empirical validation is needed for broader cases. Future work could explore adaptive sketching strategies or integration with physics-based constraints to enhance robustness.