Interest and employment in agriculture are decreasing globally. In developing countries, with growing younger populations, the number of people involved in agriculture is decreasing. In Benin, one of the main reasons that agriculture is not attractive is the lack of machinery to support farm workers. As a first step towards the development of a robotic system for maize harvesting, this study proposes a two-stage vision pipeline: YOLO-based cob detection followed by maturity classification. Using field-collected maize images, we localize cobs. For maturity grading, MobileNetV3 is the baseline, compared with CLIP zero-shot and a Prototypical Network (few-shot), which gives the best results on our dataset.

We study the approximation of nonlinear partial differential equations (PDEs) using neural operators. While traditional numerical methods suffer from high computational costs due to nonlinearities and high dimensionality, neural operators provide a promising alternative by learning mappings between infinite-dimensional function spaces. In this work, we establish a quantitative approximation theorem of neural operators for the solution operator of semilinear elliptic PDEs. Our construction is based on Banach’s fixed point theorem and Picard iteration, and demonstrates that the depth and width of the neural operator grow at most logarithmically with respect to the approximation accuracy. This result shows that exponential growth in model complexity can be avoided. Our framework is generalizable to a broader class of PDEs and offers a constructive foundation for both theoretical analysis and future empirical studies.

In contrast to forward analysis, which derives outputs from inputs, inverse analysis estimates inputs from outputs. Inverse analysis plays a crucial role in optimal design and efficient control; however, it is known to be an ill-posed problem that is difficult to solve. This study aims to construct a framework for electromagnetic inverse analysis assisted by machine learning, targeting estimation of the source current density distribution from the electric field distribution. In this paper, we report on the application of our proposed machine learning models to two-dimensional and three-dimensional electrostatic field problems, demonstrating a promising outlook for establishing this inverse analysis method.

To solve time-harmonic problems with complex symmetric matrices in electromagnetic field computations more robustly and efficiently, ten new product-type Krylov subspace (PT) methods are proposed. They are derived from the COCG or COCR methods and are evaluated through finite element analyses of the time-harmonic eddy current and high-frequency electromagnetic wave problems. As a result, in several test cases, iteration counts and the computational time can be substantially reduced. However, in high-frequency electromagnetic wave simulations, the convergence criteria required for the PT methods must be two to three orders of magnitude tighter than those required for the COCG method.

Although an autonomous mobile buoy to observe wave-height without a mooring rope has been developed, rotation of the front and rear propellers caused unintended turning during navigation. In this paper, the cause of this problem was investigated through computational fluid dynamics (CFD) analysis, and it was confirmed that the problem could be resolved by introducing a phase difference between the propellers. Conventional wave-height measurement systems, which are fixed to the seabed via mooring lines, suffer from limited coverage and high maintenance costs. Accordingly, a small autonomous mobile buoy has been developed. The buoy achieves forward propulsion through self-rotation using propellers mounted at the front and rear of a fully sealed cylindrical hull; however, unintended turning motion was observed. CFD analysis demonstrated that the interaction between the flows generated by the front- and rear-mounted propellers produces a yaw moment, causing the buoy to turn unintentionally. Furthermore, it was confirmed that the yaw moment can be reduced by introducing a phase difference between the propellers.

A pseudo-spectral large-eddy simulation (LES) code for Hall magnetohydrodynamic (MHD) turbulence is optimized by overlapping communication and computation in three-dimensional fast Fourier transforms, reducing communication overhead. The CPU-based code is ported to GPU-based systems using a minimal-intrusion approach with OpenACC, cuFFT, and limited NCCL routines, enabling rapid migration with a unified code base. LES employing a recently developed sub-grid-scale (SGS) model successfully reproduces characteristic Hall MHD turbulence features. Energy spectra of forced homogeneous and isotropic turbulence are consistent with a high-resolution reference computation and with previously reported Hall MHD turbulence spectra. Sheet-like structures of the current density and enstrophy density in our LES are also consistent with those reported in our previous Hall MHD turbulence simulations. These results confirm the applicability of the SGS model and the fidelity of the GPU implementation.