1 of 1

Continuous Physical Field Reconstruction from Sparse Observations�With the ImPACTS SciDAC5 Partnership (BER)

Scientific Achievement

We present a new method to accurately reconstruct a smooth, continuous physical fields from few sensor points. The model is based on implicit neural representations (INRs) dealing with sparse and irregular data.

Significance and Impact

Reconstructing detailed physical patterns from limited sensor information is a tough task seen in many scientific areas. We introduce a new deep neural network method to improve how we reconstruct physical data from sparse sensor readings. It separates and learns data patterns over space and time, creating a detailed, continuous model. Tests show it surpasses existing methods, providing clearer reconstructions for scientific data.

This figure shows qualitative evaluations comparing ours (MMGN) and other baseline models, which demonstrates the superiority of our work. The first row illustrates the reconstruction results of the global surface temperature while the second row shows their corresponding errors (with lighter colors depicting larger errors).

Technical Approach

Introduce a context-aware indexing mechanism that compared to standard time index (t)-based INR models, incorporates additional semantics.
Factorize target signals into a set of multiplicative basis functions and applying element-wise shift and scale transformations to the latent codes.

PI: Shinjae Yoo

SciDAC partnership: ImPACTS, Improving Projections of AMOC and Collapse Through Advanced Simulations

ASCR Program: SciDAC RAPIDS

ASCR PM: Lali Chatterjee, Kalyan Perumalla

Publication: Luo, X., Xu, W., Nadiga, B., Ren, Y., & Yoo, S. Continuous Field Reconstruction from Sparse Observations with Implicit Neural Networks. ICLR’24.

Code: https://xihaier.github.io/ICLR-2024-MMGN/

Luo, X., Xu, W., Nadiga, B., Ren, Y., & Yoo, S. (2023, October). Continuous Field Reconstruction from Sparse Observations with Implicit Neural Networks. In The Twelfth International Conference on Learning Representations.

LOCAL LAB POC:

Shinjae Yoo, Wei Xu, Xihaier Luo, Yihui Ren

TALKING POINTS:

In many scientific fields, it's difficult to accurately recreate physical environments or patterns from limited sensor data.
Often, the underlying processes that generate this data are not fully understood, making the reconstruction task even more challenging.
This study proposes a new technique using deep neural networks, specifically through implicit neural representations (INRs), to overcome these challenges.
It begins by separating the data's variability into spatial and temporal components.
Then, it uses these separated components to learn from sparsely collected, unevenly distributed data points, creating a detailed and continuous model of the physical field.
This method has been tested and shown to produce better results than existing INR methods, especially in creating high-quality reconstructions from advanced climate simulation data and high-resolution satellite images of sea surface temperatures.

METADATA:

Name of the associated awarded project: Improving Projections of AMOC and its Collapse Through advanced Simulations (ImPACTS)

PI name(s): Luke Van Roekel (LANL), Sarat Sreepathi (ORNL), Shinjae Yoo (BNL), Sri Hari Krishna Narayanan (ANL), Gokhan Danabasoglu (NCAR), Zhengyu Liu (OSU), Julie McClean (SIO), Patrick Heimbach (UT)

Name of the program manager: Lali Chatterjee, Kalyan Perumalla

CITATIONS:

Luo, X., Xu, W., Nadiga, B., Ren, Y., & Yoo, S. (2023, October). Continuous Field Reconstruction from Sparse Observations with Implicit Neural Networks. In The Twelfth International Conference on Learning Representations.
https://xihaier.github.io/ICLR-2024-MMGN/