Bayesian Data Assimilation for Low-Cost Nonlinear Simulation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional data assimilation algorithms face challenges in implementing numerical simulations due to high computational costs and the difficulty in deriving adjoint models for non-linear phenomena, especially in fields requiring large amounts of computing resources.

Innovation Solution

A novel data assimilation algorithm combining Bayesian Optimization with minimization computation of the evaluation function, eliminating the need for gradient computation and reducing computational cost, is employed to facilitate easy implementation and efficient data assimilation.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If conventional data assimilation algorithms are used to perform numerical simulations, then prediction accuracy can be improved, but computational cost increases significantly

Engineering Contradiction:
Improveprediction accuracyVSAvoidcomputational cost
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent transforms the data assimilation problem from a continuous optimization problem into a discrete Bayesian inference problem by changing the parameter representation. Instead of directly optimizing continuous parameters through gradient-based methods, the invention discretizes the parameter space and uses Bayesian probability to evaluate and update parameter candidates, significantly reducing computational burden while maintaining accuracy

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent replaces the traditional mechanical/numerical optimization system (gradient descent, adjoint models) with a probabilistic Bayesian system. This substitution eliminates the need for complex adjoint model derivations and gradient computations, using instead a framework based on probability theory and statistical inference that is computationally more efficient

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

2Measurement precision

If conventional data assimilation algorithms are used for non-linear phenomena, then prediction accuracy can be improved, but implementation difficulty increases due to adjoint model derivation

Engineering Contradiction:
Improveprediction accuracyVSAvoidimplementation difficulty
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent replaces the complex mechanical/numerical system of adjoint model derivation with a probabilistic Bayesian framework. This substitution eliminates the need for analytical adjoint models, as the Bayesian approach uses direct simulation and probability evaluation that can handle non-linear phenomena without requiring complex mathematical derivations

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent introduces Bayesian probability theory as an intermediary framework between the simulation model and the parameter optimization process. This intermediary layer simplifies the interaction by using probability distributions to represent uncertainty and guide the search for optimal parameters, avoiding the need for direct gradient computations or adjoint model derivations

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP4404213B1Data assimilation device, data assimilation method, data assimilation program, and data assimilation system
Publication Date: 2026.04.22 NAT UNIV CORP TOKYO UNIV OF AGRI & TECH
  • EP4404213B1 patent drawingFigure 1
  • EP4404213B1 patent drawingFigure 2
  • EP4404213B1 patent drawingFigure 3

AI summary

An acquisition section (101) acquires an actual measurement value of a measured change in a specific environment of a data assimilation target. A computation section (102) uses a preliminary initial state and a preliminary value of an unknown parameter that are related to the data assimilation target to perform a numerical computation of a change in the specific environment of the data assimilation target. An update section (103) computes a value of an evaluation function representing errors between the actual measurement value and a value obtained from a result of the numerical computation and corresponding to the actual measurement value, finds an acquisition function from plural combinations of values of the initial state and the unknown parameter combined with the evaluation function value, and updates values of the initial state and the unknown parameter so as to minimize a value of the evaluation function based on a value of the acquisition function. Values of the initial state and the unknown parameter related to the data assimilation target are estimated by an iteration determination section (104) repeating each processing of the computation section (102) and the update section (103) until a predetermined iteration end condition is satisfied.