Stefano De Marchi (University of Padova, INdAM Gruppo Nazionale di Calcolo Scientifico,Italy)
Title: Kernel Approximation and HPC
Abstract:
In this talk we discuss some kernel approximation techniques for working/handling huge amount of data. Indeed, kernel methods provide an elegant and principled approach to nonparametric learning, but presently could hardly be used in large scale problems, since na¨ıve implementations scale poorly with data size. We present some new approach that takes advantage of GPU architectures as well as with the curse of dimensionality phenomena occuring among the others, in numerical analysis, sampling, combinatorics, machine learning.
Koji Nishiguchi (Nagoya University, RIKEN Center for Computational Science Japan)
Title: 3D Generative AI for Structural Design: Potential and Prospects of Large Structural Models
Abstract:
Advancements in high-performance computing (HPC) have driven progress in structural design through 3D generative AI. Using the supercomputer “Fugaku”, we conducted Eulerian crash simulations on 21,998 automotive components, creating a large-scale dataset of paired 3D geometries and mechanical parameters, including height, depth, width, and impact absorption energy.
A DeepSDF-based model, developed in our laboratory, was used to generate 3D geometries from mechanical performance parameters. DeepSDF, a neural field approach, provides flexible 3D representation for arbitrary resolutions and topologies. Chamfer Distance evaluations demonstrated excellent generalization, with interpolation accuracy exceeding 99.0% and extrapolation accuracy surpassing 86.0% for height. For impact absorption energy, a nonlinear parameter, the model achieved approximately 92.5% accuracy at 10 mm deformation within the interpolation range, confirming its high precision in generative tasks.
As future work, we propose the Large Structural Model (LSM), an AI framework to democratize and automate structural design. Neural field-based scalability suggests that generative AI’s scaling laws may extend to 3D geometry generation, enabling unmatched precision.
Takahiro Katagiri (Nagoya University, Japan)
Title: Trends in Auto-Tuning Research for Quantum Annealing
Abstract:
This presentation delves into the emerging trends and critical role of auto-tuning (AT) in the realm of high-performance computing, particularly within quantum-related technologies such as quantum annealers and pseudo-quantum annealers. We will highlight the necessity of AT by exploring its application in solving complex combinatorial optimization problems, including vertex cover problems and support vector machines (SVM), through quantum annealing. By examining key performance parameters and their optimization, we aim to underscore the pivotal role of AT in enhancing quantum computing efficiency. This research has been partially funded by NTT Research, whose financial and technical support is gratefully acknowledged.
Kengo Nakajima (The University of Tokyo, Japan)
Title: Road to "AI for Science": Exploring Software Sustainability through "Couplers"
Abstract:
"Coupler" is originally a tool for coupling multiple simulation models such as atmosphere and ocean, structure and fluid. In recent years, computer systems and workloads have become more diverse, and the role of couplers in supercomputing has become more important. In this talk, we focus on the "history" of couplers and consider what software sustainability means. We briefly describe three projects, ppOpen-HPC (2011-2018), h3-Open-BDEC (2019-2024), and JHPC-quantum (2023-2028), and will introduce how couplers have evolved and what role they have been playing in supercomputing.
Katsuhisa Ozaki (Shibaura Institute of Technology, Japan)
Title: Fast and accurate algorithm for matrix multiplication using fused multiply-add
Abstract:
We introduce a new algorithm for high-precision computations of matrix multiplications. While hardware-supported floating-point operations are fast, they suffer from rounding errors due to their finite precision. When the accuracy of the computed results is not satisfactory, high-precision computation may be considered.
One option is using multi-precision arithmetic, such as MPFR, but if extending the range of the exponent part is unnecessary, an alternative is to represent numbers as the sum of floating-point numbers and perform operations on those sums. Examples include pair arithmetic by Lange and Rump and double-word arithmetic by Bailey. In this study, we propose an algorithm that leverages this structure for fused multiply-add operations and applies it to matrix multiplication. As a result, we have designed a computation method that is less costly than pair arithmetic or double-word arithmetic, allowing for a slight degradation in accuracy. Finally, we demonstrate the performance of the proposed method through numerical experiments. This is joint work with Dr. Toru Koizumi.
Koki Masui (Osaka University , Japan)
Title: A preconditioner for COCG method in large-scale problems
Abstract:
Iterative methods for large-scale electromagnetic field analyses often suffer from slow convergence. IC (Incomplete Cholesky) preconditioning with fill-in is used to improve the convergence. However, it can lead to memory shortage in large-scale problems. In this paper, we propose a preconditioner that improves the convergence and reduces memory usage. We show some numerical examples and evaluate the performance.
Yuki Uchino (RIKEN)
Title: High-Performance Eigensolver Combining EigenExa and Iterative Refinement
Abstract:
This study proposes a high-performance and reliable eigensolver via mixed-precision arithmetic between ordinary and highly-accurate precisions. Eigenvalue decomposition is ubiquitous in simulations. Various eigensolvers for computing approximations have been developed thus far. If eigenvalues are narrowly clustered, the computation of eigenvectors may be ill-posed. Thus, the computed eigenpairs may not be sufficiently accurate and lack reliability. In this study, we introduce mixed-precision iterative refinement methods to improve the accuracy of eigenvectors obtained using numerical methods. This approach contributes to obtaining sufficiently accurate results without arbitrary precision eigensolvers. We construct a high-performance and reliable eigensolver by combining the iterative refinement methods and EigenExa, a modern high-performance solver for large-scale and highly parallel computations. Numerical experiment results demonstrate the accuracy of the results and performance benchmark of the proposed approach.
Hiroyuki Takizawa (Tohoku University, Japan)
Title: Advanced resource management for urgent job execution in Connected Supercomputing
Abstract:
Connected Supercomputing" is to enable a supercomputing system in academic use to timely respond to external factors connected to that system. For Connected Supercomputing, we must develop new operation policies and system software for dynamic and adaptive resource management. Thus, the ExpressHPC project has started in 2024 to offer an expressway to urgent allocation of computing resources from multiple datacenters to support time-critical simulations. This talk will introduce the ExpressHPC project and then discuss the technical challenges and potential opportunities of research collaboration.
Teruo Tanaka (Kogakuin University, Japan)
Title: Search for optimal combinations of performance parameters and development of tools
Abstract:
We have been studying search methods to find optimal combinations of possible values of performance parameters. The basic idea is to use a discrete-type spline function called d-Spline as the fitting function. The target application is run while automatically changing combinations of possible values of the performance parameters. The obtained evaluation value is added and the fitting function d-Spline is updated. Initially, this function was used to estimate a single performance parameter, but later it could be applied to multiple performance parameters. As an application, it can be applied to hyperparameters in machine learning. It is parallelized to handle large numbers of computations in machine learning applications. Currently, this performance parameter search is released as a beta version of the tool DSICE.
Toshiyuki Imamura (RIKEN, Japan)
Title: mX_Real, yet another multi-component floating point package
Abstract:
Several multi-component formats for higher precision are known as double-double (DD), quadruple-double (QD), or doubled-float in the early GPGPU era. Ozaki and Imamura investigate a slightly less accurate but efficient algorithm that offers a new interface class for C++, mX_Real. The key idea is to eliminate regular normalization in every operation and flexible mechanisms to switch non-normalized or normalized modes with the user's descriptions. Also, taking advantage of modern programming languages, mX_Real offers users a flexible multi-component floating point number format and is currently being released for practical use. In this presentation, we report an overview of mX_Real and its roadmap for 2025.
Kenji Ono (Kyushu University, Japan)
Title: Solving Large-Scale Sparse Linear Systems Using Tensor Train Decomposition
Abstract:
In the post-Moore era, reducing memory usage and computational cost has become a critical challenge for achieving high-speed computation. Tensor Train Decomposition (TTD) is a technique that compresses data to reduce memory usage while allowing computations to be performed in the compressed form. This feature presents the potential to lower the B/F ratio (bytes-to-flops ratio) in iterative solvers for large-scale sparse matrices, which typically demand high memory bandwidth. In this presentation, we report the results of a study conducted on sparse matrices derived from the finite difference formulation of the Laplace equation. By comparing with exact solutions, we evaluate the computational accuracy, memory usage, and computation time as functions of the compression parameters.
Yu-Chen Shu, Jehn-Ruey Jiang and Chao-Yi Lin
Title: Comparison study for solving quadratic unconstrained binary optimization by classical and digital annealer
Abstract:
In this talk, I will delve into the combinatorial optimization problems formulated in the quadratic unconstrained binary optimization (QUBO) framework and solve them by the classical and digital annealers. I will present a comparative analysis of digital annealers versus traditional annealers, focusing on their performance across a range of classical NP-hard problems. Additionally, I will show some recent results for the one-way-one-hot constraint, highlight several applications where digital annealers have been successfully employed, showcasing the significant results achieved.
Heng-Chuan Kan
Title: Accelerating Adaptive Neural Particle Simulations with Transfer Learning
Abstract:
The Adaptive Neural Particle Method (NPM-LA) is a physics-informed neural network (PINN)-based particle method designed for simulating free surface flows with high accuracy. Despite its advantages, NPM-LA faces significant computational inefficiencies due to its reliance on a discrete time model (DTM) for time evolution, requiring sequential training at each timestep. To address these limitations, this study explores the integration of transfer learning to enhance the efficiency of NPM-LA. Transfer learning, a technique that applies pre-trained models to related tasks, has the potential to significantly reduce computational time and resources.
The proposed approach is particularly beneficial for simulations involving time-dependent fluid dynamics, where efficiency and accuracy are paramount. The results indicate that integrating transfer learning into NPM-LA not only accelerates computation but also optimizes resource utilization. Additionally, the method offers a scalable framework for extending machine learning applications to other particle-based fluid simulations. The successful implementation of transfer learning in NPM-LA represents a significant advancement in computational fluid dynamics, enabling faster and more efficient simulations of complex fluid phenomena. This study highlights the potential for machine learning-driven approaches to overcome computational bottlenecks and enhance predictive modeling capabilities in fluid mechanics. Future work will focus on refining hyperparameter selection, exploring additional test cases, and evaluating the approach's adaptability to diverse computational fluid dynamics scenarios.
Y.Nishimura
Title: Kinetic simulation of charged particle systems: Particle-in-Cell and N-body approaches
Abstract:
To analyze kinetic behavior of charged particle systems one can think of two approaches. The most commonly employed method (for industrial, fusion, and space plasmas) is the Particle-in-Cell (PIC) approach,[1] which is based on the idea of principle of locality. In PIC, we solve for Lagrangian variables (by equations of motion) and field quantities (by Maxwell equations) alternatively. It is known that the larger the scale of the PIC type calculation, the more cpu-time is spent for solving the field quantities. [2] The basic idea behind the N -body simulation is action at a distance.[3] Without introducing macroscopic fields, the Coulomb forces between charged particles are computed directly. The forces are superposed. Note that unlike the gravitational problems where all the forces are attractive, in our problem, the forces can be attractive or repulsive depending on the pair of charges. As a result, Debye shielding is inherent, which allows us to compute Coulomb forces from a finite number of particles. Plasma N-body simulations are less laborious compared with gravitational N-body simulations. The number of Coulomb interactions that need to be computed in each time step is expected to scale as ~ Λ N, where Λ is the "plasma parameter" (the number of charged particles within the Debye sphere) in place of N (N-1) or N log(N). Message Passing Interface based parallel computing applied to both the approaches are discussed.
[1] C.W.Huang, Y.C.Chen, and Y.Nishimura, IEEE Transactions on Plasma Science 43, 675 (2015).
[2] Y.Nishimura et. al, Journal of Computational Physics 214, 657 (2006).
[3] C.P.Wang and Y.Nishimura, IEEE Transactions on Plasma Science 47, 1196 (2019).
Yi-Ju Chou, Yu-Jen. Chang, and Hsuan-Yu Huang
Title: A multiscale computational framework using active learning to model complex suspension flows
Abstract:
We devise a novel computational framework to achieve efficient multiscale modeling for suspension flows. The modeling framework comprises a particle-resolving direct numerical simulation for microscopic computation, a single-fluid continuum model to capture the bulk flow behavior at the macroscale, and a Gaussian process regression that connects the two modeling components. The microscopic calculation simulates detailed flow fields around individual suspended particles. Along with a coarse-grained method, we are able to obtain the mean volume fractions and stresses for discrete particles, which are then further used to calculate the rheological properties of the bulk flow fields at the macroscale under specific forcing conditions. Using Gaussian process regression, a few data points from the microscopic calculations can be further interpolated/extrapolated to form complete constitutive relationships that cover the whole forcing range in macroscopic calculations (i.e., continuum modeling). Moreover, via the uncertainties returned from the Gaussian process regression, the resulting active learning strategy using the Gaussian process regression automatically decides on the required the microscopic calculations given by a specified tolerance. As a result, complex suspension problems can be efficiently solved by the macroscopic continuum model to the desired accuracy with the minimum computational effort (i.e., minimum number of microscopic runs).
Te-Yao Chiu (Department of Mechanical Engineering, National Central University, Taiwan)
Title: Data-Driven Analysis and Reduced-Order Modeling of Flow Structures' Impact on External Flow Fields
Abstract:
This study consists of two parts. The first part investigates the contributions of coherent structures to the aerodynamic forces exerted on a NACA0012 airfoil at angles of attack of 5, 10 degrees. Utilizing Proper Orthogonal Decomposition (POD) in conjunction with vorticity force analysis, we assessed their contributions to lift and drag forces at a chord-based Reynolds number of 50,000. At the smallest AoA, the primary source of time-varying aerodynamic forces arises from the detachment of the spanwise vortex at the reattachment point. In this case, the zeroth POD mode (mean flow) has the dominant contribution to the total force, with contributions from the first few non-zero POD modes being indistinct. As AoA increases to 10 degrees, the first POD mode corresponds to an apparent vortical structure located at the second half section of the chord. The clockwise rotation of this vortical structure leads to a strong re-entrant flow on the airfoil’s suction side, resulting in a positive drag contribution due to the intense shear near the boundary. The second POD mode in the case with AoA = 10 degrees corresponds to a trailing edge vortex (TEV), which counteracts the vortical structure of the first mode, leading to a decrease in drag force.
The second part aims to develop a reduced-order model based on data-driven methods. The core concept of this model is to extract the main features and temporal coefficients of the flow field data using proper orthogonal decomposition. Then, a high-order dynamic mode decomposition is used to further simplify the temporal coefficients, and Gaussian process regression is employed to establish the relationship between temporal coefficients under different flow parameters. Finally, by interpolating the temporal coefficients corresponding to new flow parameters and combining them with the modes, we can obtain flow field results different from those in the training dataset. The advantage of this study lies in significantly reducing computational requirements while retaining the key features of the system, thereby accelerating parameter adjustment and optimization process and making them more adaptable to various application scenarios.
Che-Wei Tsao, Yun-Hsin Lin (National Chung Hsing University, Taiwan)
Title: 6 DOF FSI Approach in Hierarchical Cartesian Mesh Framework by Immersed Boundary Method
Abstract:
This study presents a high-fidelity computational fluid dynamics (CFD) framework that integrates six degrees of freedom (6DOF) motion equations with the immersed boundary method within a hierarchical Cartesian mesh framework. Traditional CFD simulations often assume fixed objects, neglecting fluid-structure interactions. To address this, the proposed approach allows bidirectional coupling between fluids and moving objects, capturing both translational and rotational dynamics without requiring complex grid deformations. The framework is validated through multiple test cases, including a propeller model generating thrust under rotation and a fan blade model rotating in response to an applied wind field. Additionally, a two-way coupled simulation of a swimming shark is conducted. A motion equation governs the shark’s self-propelled undulatory movements, which induce thrust and other hydrodynamic forces, demonstrating the method's ability to model biologically inspired locomotion. The combination of the hierarchical Cartesian mesh framework and the immersed boundary method provides an efficient and accurate means of simulating complex fluid-structure interactions. This methodology has broad applications in aerospace engineering, bio-inspired propulsion, and hydrodynamics. Future work will focus on refining the motion models and extending the approach to more complex real-world scenarios.
BingZe Lu, Richard Tsai
Title: CAVEATS IN DYNAMICAL SYSTEMS LEARNING
Abstract:
In this talk, I will introduce the ODEnet, which integrates conventional numerical methods to define loss functions. One of its applications focuses on identifying the underlying dynamical system from reachable observation snapshots. Considering an autonomous differential system dx/dt = f(x), the ODEnet approximates the unknown function f(x) by a trainable neural network and minimizes the difference of predictions generated by using numerical methods to the observation data at a given sampling time stamp. In this talk, I will focus on discussing the case of learning linear dynamical systems, which are either conservative or dissipative.
The first part of my talk will address structural issues, including rotation and trends in the learned dynamical systems, using one-step or multi-step methods. While the loss function can be ideally minimized, the learned results may not accurately reflect the system's true behavior. I will present our findings and conclude this section by proposing criteria for selecting numerical integrators that retain the unknown dynamics' structure.
The second part will analyze the impact of noisy data on the learning outcomes. I will demonstrate how different noise levels distort the results obtained from noisy data compared to those inferred from noise-free data.
Matthew R. Smith (Department of Mechanical Engineering, National Cheng-Kung University, Taiwan)
Title: SCALAR VECTOR EXTENSION (SVE) INTRINSICS FOR COMPUTATIONAL FLUID DYNAMICS ON NEXT GENERATION ARM ARCHITECTURES IN AWS
Abstract:
Computational Fluid Dynamics (CFD) has become a mainstay of modern engineering design and analysis, though its adoption is hindered by its high computational expense. To overcome this, the application of parallel computation techniques is a common strategy in decreasing the computational time. In the past, this has been done effectively on large scale distributed memory systems, shared memory systems and via application of Graphical Processing Units (GPUs). In the past, low level optimization of parallel CFD codes using AVX SIMD vectorization on x86 systems has also been used to great effect.
Presented in this work is the application of a parallel Finite Volume Method (FVM) to simulation of shallow water flow and compressible gas flow using the SHLL algorithm on ARM cores using AWS’s new Graviton4 (Neoverse v2) CPUs. These architectures differ from vectorization on conventional x86 as they employ the new Scalar Vector Extension (SVE) instruction set which allows application of larger registers (up to 2048 bits), greatly increasing local parallelism and computational efficiency. These solvers are executed on AWS resources defined and created using AWS’s Cloud Development Kit (CDK). This talk will discuss the use of AWS CDK for rapid acquisition of resources in AWS, the SHLL solver and application of SVE intrinsic functions together with OpenMP for large scale parallel computation of hyperbolic flows.
Torbjörn E. M. Nordling (Department of Mechanical Engineering, National Cheng-Kung University, Taiwan)
Title: Energy-Efficient Inference in Deep Neural Networks: Trading Accuracy for Reduced Compute through Model Compression
Abstract:
Deep neural networks (DNNs) are increasingly deployed in resource-constrained environments where energy efficiency is paramount. Achieving real-time inference on edge devices often requires a balance between computational cost and predictive accuracy. This talk explores how model compression through Knowledge Distillation (KD) can intentionally forgo accuracy to significantly reduce compute and energy usage.
I take a look at eight published applications of KD across five well-known datasets to examine the trade-offs between model compression and accuracy in real-world scenarios. To overcome the lack of data on different compression factors, we have in my lab conducted systematic experiments reducing the size of a model up to 1000 times.
We trained a MobileNet-V2 teacher model on CIFAR-10 to achieve a testing accuracy of 93.91%. The compression experiments were done on a student model with an adopted MobileNet-V2 architecture involving removal of the last layers before adding max pooling and a new classification layer to achive compression factors ranging from 2.19 to 1121.71. The students were trained independently and using KD with integrated gradients (IG) for data augmentation.
Our experiment demonstrates that the accuracy can be maintained above 90% while compressing the number of model parameters by a factor of up to 12.5, but then it drops off below 80% at 100 and 60% at 1000. The model can be compressed by a factor of 4.12, reducing its computational footprint while maintaining 92.45% accuracy, only 1.46 percentage points lower than the original model. This presentation discusses these results and provide guidance on trading accuracy for efficiency, thus enabling AI deployment in mobile, embedded, and cloud-edge computing.
By systematically quantifying the impact of compression on accuracy or other performance metrics deep learning models can be adapted for low-power environments without compromising essential functionality. The best way of saving energy and time is by avoiding unnecessary computations. Past optimisation of algorithms is in the deep learning era replaced by model compression.
Makoto Morishita (D3, Nagoya University, Japan)
Title: Adaptation of Auto-Tuning to Quantum Annealers
Abstract:
In this talk, we will discuss the application of automatic tuning (AT) to the algorithm to a Coherent Ising Machine (CIM) and an inspired quantum annealer. The CACm (Chaotic Amplitude Control with momentum) used for ground state search of the Ising Hamiltonian for annealing with CIM. In addition, we are focusing on CMOS Annealing Machine by Hitachi Ltd. for the inspired quantum annealer. Optuna, developed at Preferred Networks, Inc. , was used as the AT software in this research. The CIM-CACm has six performance parameters. We are now trying a method to adapt Optuna to improve performance of CIM-CACm more efficiently. On the other hand, we have developed the AT function to the CMOS Annealing Machine for solving Maximal Covering Location Problem (MCLP). The target is two parameters: a link strength and a weight of QUBO formula. AT effect by Optuna for MCLP was evaluated with several problem sizes. In this presentation, we will show the state-of-the art results. This research has been partially funded by NTT Research, whose financial and technical support is gratefully acknowledged.
Hua-Ching Chang (National Chung Hsing University / National Center for High-Performance Computing,Taiwan)
Title: MULTI-GPU DIRECT NUMERICAL SIMULATION OF HIGH RAYLEIGH NUMBER NATURAL CONVECTION OVER A HEATED CIRCULAR PLATE
Abstract:
Natural convection plays a critical role in fluid heat transfer and is widely observed in both natural phenomena and engineering applications. This study investigates the natural convection of a large circular heating plate under high Rayleigh number conditions (Ra = 4.2 × 10^8) using computational fluid dynamics (CFD). Direct numerical simulation (DNS) integrated with a multi-node multi-GPU computational framework improves computational efficiency and significantly reduces computation time. Non-reflective boundary conditions are implemented to minimize numerical pressure wave reflections, effectively reducing the need for external grid. Simulations are conducted for heating plates with varying diameters to explore the development of natural convection.
The simulation results reveal that at extremely high Rayleigh numbers, the large circular heated plate undergoes flow transitions near its edges, resembling those observed in forced convection over horizontal heated plates. These transitions progress from laminar to transitional and ultimately to turbulent flow. In the laminar region, cold air is drawn inward from the periphery, while the transitional region gradually develops upward flow. In the turbulent region, significant energy is released from the heated plate, giving rise to coherent turbulent structures.
Shubham (D1, Tohoku University, Japan)
Title: Utilizing Hardware Performance Counters to Forecast Workload Interference in Vector Supercomputer
Abstract:
High-performance computing (HPC) systems, such as the SX-Aurora TSUBASA (SX-AT), integrate heterogeneous architectures that combine diverse processor types, presenting both opportunities and challenges in resource optimization. This study explores workload interference within SX-AT systems, focusing on resource contention between Vector Hosts (VHs) and Vector Engines (VEs). Through empirical analysis, we identify critical factors contributing to performance degradation, including cache and memory bandwidth contention when workloads with varying computational demands share resources. To address these challenges, we propose a predictive model that utilizes hardware performance counters (HCs) and machine learning (ML) algorithms to classify and predict workload interference. The model accurately forecasts performance degradation, offering insights to enhance job scheduling and resource allocation strategies. This research underscores the importance of adaptive resource management in heterogeneous supercomputing environments and lays the groundwork for future advancements in system efficiency.
Ryota Koda (M2. Tohoku University, Japan)
Title: Development of a Real-Time 3D X-ray Ptychography Workflow Using Surrogate Models
Abstract:
X-ray ptychography is a non-destructive imaging method with high spatial resolution applied to various research fields, such as materials science and electronics. X-ray ptychography reconstructs a sample image using iterative phase retrieval from its diffraction patterns. However, iterative phase retrieval is time-consuming, and thus real-time reconstruction is challenging.
One approach to accelerating iterative phase retrieval involves using a deep learning based surrogate model to predict the sample. By performing iterative phase retrieval using the model predictions as initial estimates, it is possible to achieve faster reconstruction compared to conventional iterative phase retrieval algorithms. However, discussions on the performance of these methods often neglect the time required for training and dataset preparation. Therefore, conventional iterative methods might rather be faster if these times are considered.
The objective of this study is to develop a real-time 3D X-ray ptychography workflow, including model training and dataset creation. For this purpose, we propose a method to create a sample-specific surrogate model on-the-fly from a subset of observed data and to perform iterative phase retrieval using the predictions of the model.
The proposed method is up to 3.2 times faster than conventional iterative phase retrieval, including training dataset creation time and surrogate model training time. Moreover, it was found that preparing a surrogate model with moderate accuracy in a shorter time reduced the workflow execution time compared to spending more time preparing a highly accurate surrogate model. This result indicates that it is important to find a good trade-off point between model accuracy and model training time.
Kanae Shiragami (M1, Kyushu University, Japan)
Title: Estimation of wind turbine wake models by Genetic Programming
Abstract:
Genetic programming (GP) is an algorithm that searches for optimal solutions through iterative evolution and selection of individuals. GP can be applied to equation discovery as a method for solving symbolic regression problems. By referring to existing equations and terms, we extend the definitions of “function” and “terminal” in GP to enable the estimation of various forms of equations. This allows us to search for representations of unknown equations and models for phenomena within a large candidate space. In this presentation, we report our efforts to estimate a wind turbine model and an engineering wake model using high-precision simulation data from the area surrounding a wind turbine. Our final goal is to develop surrogate models of wind turbine wakes for application in wind farms in the future.
Yu-Tan Wu (M2, National Central University, Taiwan)
Title: A Hybrid Deep Learning and Data Assimilation Approach on Enhancing Accuracy of Foreign Currency Exchange Rate Forecast
Abstract:
Accurate forecasting methods remain a fundamental challenge in many scientific fields, such as meteorology, oceanography, and nonlinear dynamical systems. Traditional methods based on physical models often fail due to uncertainties in model parameters, while deep learning methods struggle with long-term generalization. In this work, we demonstrate a hybrid approach that integrates LSTM (Long-Short Term Memory) networks with data assimilation techniques, specifically the Extended Kalman filter (EKF), to improve forecast accuracy. The LSTM network is trained on historical data to capture long-term patterns, while the Extended Kalman Filter (EKF) corrects the initial conditions for forecasting using observational data. We evaluate our approach through foreign currency exchange rate forecasting and compare its performance against pure deep learning methods. This study highlights the potential of combining deep learning and data assimilation to improve forecast results.