1 of 1

Explorable INR for Ensemble Simulation

With the BER ImPACTS SciDAC5 Partnership program

Scientific Achievement

Implicit Neural Representation (INR) is a popular way to reduce scientific data. We develop an INR-based surrogate model for ensemble simulations and explore both the spatial and simulation parameter domain through our INR-Surrogate.

Significance and Impact

Our INR-Surrogate can accurately predict ensemble simulations and enable random access to the values at a given location and parameter setting without the need for full-scale field data reconstruction. Based on INR-Surrogate, we develop an efficient technique to gain an overview of ensemble simulations and identify parameter setting for desired physical attribute distribution.

INR-Surrogate employs feature grids as encoder for spatial and simulation parameter input. The feature vectors are concatenated to form ensemble feature and decoded by MLP to gain the result.

Blue path: Given a simulation parameter setting and coordinates in the ocean, INR-Surrogate predicts the temperature of the ocean. Light yellow for high temperature; green for mid temperature, and blue for low temperature

Orange path: Given a range of simulation parameters, we can compute the variance field, which is the variance of each coordinate in the ocean. Purple for large variance; blue for small variance.

Technical Approach

INR-Surrogate is composed of feature grid encoder and MLP decoder.
Employ uncertainty propagation to efficiently gain physical attribute distribution, i.e. mean and variance, of ensemble simulations.
Utilize gradient descent to identify possible parameter settings for a given distribution.

PI: Han-Wei Shen; Ohio State University

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

ASCR Program: SciDAC RAPIDS2

Publication: Explorable INR: An Implicit Neural Representation for Ensemble Simulation Enabling Efficient Spatial and Parameter Exploration, Yi-Tang Chen, Haoyu Li, Neng Shi, Han-Wei Shen, submitted for publication

LOCAL LAB POC:

TALKING POINTS:

The proposed INR-Surrogate learns the mapping from coordinate and simulation parameter to scalar value. With this mapping, we can do range query on either spatial coordinate or simulation parameter. Through the range query, our INR-Surrogate allows the exploration on both the spatial and parameter domain.
INR-Surrogate helps scientists explore spatial domain by efficiently computing the physical attribute distribution on a given simulation parameter range. For example, given the parameter range for ocean simulation, we can calculate the mean and variance of the ocean temperature for any location in the ocean.
INR-Surrogate helps scientists explore parameter domain by identifying the possible parameter setting for a desired physical attribute distribution. For instance, given an ocean temperature distribution which is related to hurricane, we can find the possible parameters that match this distribution. Scientists can further investigate these parameters.
In conclusion, INR-Surrogate not only accurately predict the simulation result, but also allows exploration on the ensemble simulations.

METADATA:

Name of the associated awarded project:

PI name(s): Han-Wei Shen (OSU)

Name of the program manager:

BACKGROUND AND CONTEXT INFORMATION:

Scientists run ensemble simulations to investigate complex physical system, such as ocean, but it takes long time to run simulation. Thus, researchers develop surrogate models to efficiently predict the simulation results.
To understand the physical system, scientists may be interested in the physical attribute distribution, such as temperature, on the parameter range.
Existing surrogate model can only achieve this goal by running the surrogate model multiple times to gain the approximate distribution. To gain more accurate distribution, we need to run surrogate model on hundreds of parameter settings. This is computation and memory intense.
To solve this problem, we tailor the INR to develop the explorable ensemble surrogate model, which enables range query for parameter range. That means we can calculate the distribution without dense sampling on parameter ranges.