Image by cybermagician/Shutterstock

Numerical simulations in physics often require estimating a multitude of parameters, making the process computationally expensive and complex. Researchers at University of Tsukuba have introduced a new method called the multiparameter eigenvalue-problem emulator, enabling reliable predictions based directly on relationships among known data by eliminating the need for parameter estimation. This innovation considerably reduces computational costs and enables systematic quantification of predictive uncertainty.

Tsukuba, Japan—Calibrating theoretical models with experimental data is a common practice in physics for predicting previously unobserved phenomena. However, real-world theoretical models are often highly complex, involving numerous numerical quantities, known as "parameters," that cannot be directly measured. Researchers must estimate these parameters to compute other observables. This is a process that is computationally demanding and fraught with remarkable challenges in assessing how uncertainties in the parameters affect final predictions.

This study presents a novel fast surrogate model based on a mathematical framework known as the multiparameter eigenvalue-problem emulator. This model directly predicts unknown observables based on relationships among known data, without the need to introduce or estimate parameters.

Validation against a traditional model showed that the proposed method can reliably reproduce complex behaviors that are challenging for conventional computational approaches. In addition, when applied to a nuclear physics problem, namely, predicting the energies of oxygen isotopes, the method produced probability distributions closely aligned with experimental observations. This framework supports rapid execution of large-scale computations and enables systematic quantification of predictive uncertainty.

These findings pave the way for efficient predictions of physical phenomena by removing the need for parameter estimation, a major computational bottleneck in physics research. The proposed method is anticipated to have broad applicability across various disciplines, including astrophysics and materials science.

###
This work is in part supported by JST ERATO Grant No. JPMJER2304, Japan. This work is also in part supported by the Multidisciplinary Cooperative Research Program in CCS, University of Tsukuba.

Original Paper

Title of original paper:

An Efficient Learning Method to Connect Observables

Journal:

Physical Review Letters

DOI:

10.1103/33q9-76qp

Correspondence

Researcher Hang Yu
Assistant Professor MIYAGI Takayuki
Center for Computational Sciences, University of Tsukuba

Related Link

Center for Computational Sciences