Explore the vast range of titles available.

Uncertainty quantification (UQ) is important to machine learning (ML) force fields to assess the level of confidence during prediction, as ML models are not inherently physical and can therefore yield catastrophically incorrect predictions. Established a-posteriori UQ methods, including ensemble methods, the dropout method, the delta method, and various heuristic distance metrics, have limitations such as being computationally challenging for large models due to model re-training. In addition, the uncertainty estimates are often not rigorously calibrated. In this work, we propose combining the distribution-free UQ method, known as conformal prediction (CP), with the distances in the neural network’s latent space to estimate the uncertainty of energies predicted by neural network force fields. We evaluate this method (CP+latent) along with other UQ methods on two essential aspects, calibration, and sharpness, and find this method to be both calibrated and sharp under the assumption of independent and identically-distributed (i.i.d.) data. We show that the method is relatively insensitive to hyperparameters selected, and test the limitations of the method when the i.i.d. assumption is violated. Finally, we demonstrate that this method can be readily applied to trained neural network force fields with traditional and graph neural network architectures to obtain estimates of uncertainty with low computational costs on a training dataset of 1 million images to showcase its scalability and portability. Incorporating the CP method with latent distances offers a calibrated, sharp and efficient strategy to estimate the uncertainty of neural network force fields. In addition, the CP approach can also function as a promising strategy for calibrating uncertainty estimated by other approaches.

In this work, we evaluate the capability of Neural Network-based force fields, particularly NeuralIL ( J Chem. Inf. Model. 62 , 88-101, 2021 ), to simulate complex charged fluids. We focus on how this novel force field can address several pathological deficiencies of classical force fields for such systems. First, we review the capability of NeuralIL to replicate the molecular structures of the system. Then, we analyze the structural and dynamic properties, showing that weak hydrogen bonds are significantly better predicted and that their dynamics are not hindered by the absence of polarization of the electronic densities as seen in classical force fields. Finally, we analyze the capability of NeuralIL to model systems with proton transfer reactions, demonstrating its ability to find and reproduce the reactions that take place within the system. Moreover, we validate our results by comparing them with previous predictions of the equilibrium coefficient for the same system, finding a strong agreement.

Uncertainty quantification (UQ) is important to machine learning (ML) force fields to assess the level of confidence during prediction, as ML models are not inherently physical and can therefore yield catastrophically incorrect predictions. Established a-posteriori UQ methods, including ensemble methods, the dropout method, the delta method, and various heuristic distance metrics, have limitations such as being computationally challenging for large models due to model re-training. In addition, the uncertainty estimates are often not rigorously calibrated. In this work, we propose combining the distribution-free UQ method, known as conformal prediction (CP), with the distances in the neural network's latent space to estimate the uncertainty of energies predicted by neural network force fields. We evaluate this method (CP+latent) along with other UQ methods on two essential aspects, calibration, and sharpness, and find this method to be both calibrated and sharp under the assumption of independent and identically-distributed (i.i.d.) data. We show that the method is relatively insensitive to hyperparameters selected, and test the limitations of the method when the i.i.d. assumption is violated. Finally, we demonstrate that this method can be readily applied to trained neural network force fields with traditional and graph neural network architectures to obtain estimates of uncertainty with low computational costs on a training dataset of 1 million images to showcase its scalability and portability. Incorporating the CP method with latent distances offers a calibrated, sharp and efficient strategy to estimate the uncertainty of neural network force fields. In addition, the CP approach can also function as a promising strategy for calibrating uncertainty estimated by other approaches.

One of the major challenges in the development of universal machine learning interatomic potentials is accurately reproducing phonon properties. This issue appears to arise from the limitations of available datasets rather than the models themselves. To address this, we develop an extensive dataset of phonon calculations using density-functional perturbation theory (DFPT). We then show how this dataset can be used to train neural-network force fields, by implementing the training and the prediction of force constants in periodic crystals. This approach improves the quality of phonon properties prediction while reducing the number of structures needed for neural network training. We demonstrate the efficiency of this method using two examples of ternary phase diagrams: Ti–Nb–Ta and Li–B–C. In both cases, neural network predictions for the energy and forces show a considerable improvement, while phonon properties are predicted with high precision for all structures across the entire phase diagrams.

Riverbed material has multiple functions in river ecosystems, such as habitats, feeding grounds, spawning grounds, and shelters for aquatic organisms, and particle size of riverbed material reflects the tractive force of the channel flow. Therefore, regular surveys of riverbed material are conducted for environmental protection and river flood control projects. The field method is the most conventional riverbed material survey. However, conventional surveys of particle size of riverbed material require much labor, time, and cost to collect material on site. Furthermore, its spatial representativeness is also a problem because of the limited survey area against a wide riverbank. As a further solution to these problems, in this study, we tried an automatic classification of riverbed conditions using aerial photography with an unmanned aerial vehicle (UAV) and image recognition with artificial intelligence (AI) to improve survey efficiency. Due to using AI for image processing, a large number of images can be handled regardless of whether they are of fine or coarse particles. We tried a classification of aerial riverbed images that have the difference of particle size characteristics with a convolutional neural network (CNN). GoogLeNet, Alexnet, VGG-16 and ResNet, the common pre-trained networks, were retrained to perform the new task with the 70 riverbed images using transfer learning. Among the networks tested, GoogleNet showed the best performance for this study. The overall accuracy of the image classification reached 95.4%. On the other hand, it was supposed that shadows of the gravels caused the error of the classification. The network retrained with the images taken in the uniform temporal period gives higher accuracy for classifying the images taken in the same period as the training data. The results suggest the potential of evaluating riverbed materials using aerial photography with UAV and image recognition with CNN.

The analysis of images using deep learning techniques makes it possible to detect anomalous or dangerous situations in different fields of application. This work aims to ensure the correct configuration of landing gear during aircraft landings. In contrast with other works, the small object detection problem is solved using background subtraction technique, and subsequently feeding it to our proposed convolutional neural network to automatically classify the position of the landing gear. This work also develops a new database that combines synthetic and real images, generated from exclusive fighter landing manoeuvres performed by a real test pilot. The obtained model, trained with synthetic data and tested with real images, presents a 0.9981 of accuracy. The result is a functional system, tested against real images and endowed with “early warning” capability as it is able to detect the position of an aircraft’s landing gear in advance and prevent catastrophic accidents.

Wind power can significantly contribute to the transition from fossil fuels to renewable energies. Airborne Wind Energy (AWE) technology is one of the approaches to tapping the power of high-altitude wind. The main purpose of a ground-based kite power system is to estimate the tether force for autonomous operations. The tether force of a particular kite depends on the wind velocity and the kite’s orientation to the wind vector in the figure-eight trajectory. In this paper, we present an experimental measurement of the pulling force of an Airush Lithium 12 m2 kite with a constant tether length of 24 m in a coastal region. We obtain the position and orientation data of the kite from the sensors mounted on the kite. The flight dynamics of the kite are studied using multiple field tests under steady and turbulent wind conditions. We propose a physical model (PM) using Artificial Neural Network (ANN) and Long Short-Term Memory (LSTM) deep neural network algorithms to estimate the tether force in the experimental validation. The performance study using the root mean square error (RMSE) method shows that the LSTM model performs better, with overall error values of 126 N and 168 N under steady and turbulent wind conditions.

In this study, a neural network and a multi-objective genetic algorithm were used to optimize the geometric parameters of segmented thermoelectric generators (TEGs) with trapezoidal legs, including the cold end width of thermoelectric (TE) legs (Wc), the ratios of cold-segmented length to the total lengths of the n- and p-legs (Sn,c and Sp,c), and the width ratios of the TE legs between the hot end and the cold end of the n- and p-legs (Kn and Kp). First, a neural network with high prediction accuracy was trained based on 5000 sets of parameters and the corresponding output power values of the TEGs obtained from finite element simulations. Then, based on the trained neural network, the multi-objective genetic algorithm was applied to optimize the geometric parameters of the segmented TEGs with the objectives of maximizing the output power (P) and minimizing the semiconductor volume (V). The optimal geometric parameters for different semiconductor volumes were obtained, and their variations were analyzed. The results indicated that the optimal Sn,c, Sp,c, Kn, and Kp remained almost unchanged when V increased from 52.8 to 216.2 mm3 for different semiconductor volumes. This work provides practical guidance for the design of segmented TEGs with trapezoidal legs.

Electronic skin is an important means through which robots can obtain external information. A novel flexible tactile sensor capable of simultaneously detecting the contact position and force was proposed in this paper. The tactile sensor had a three-layer structure. The upper layer was a specially designed conductive film based on indium-tin oxide polyethylene terephthalate (ITO-PET), which could be used for detecting contact position. The intermediate layer was a piezoresistive film used as the force-sensitive element. The lower layer was made of fully conductive material such as aluminum foil and was used only for signal output. In order to solve the inconsistencies and nonlinearity of the piezoresistive properties for large areas, a Radial Basis Function (RBF) neural network was used. This includes input, hidden, and output layers. The input layer has three nodes representing position coordinates, X, Y, and resistor, R. The output layer has one node representing force, F. A sensor sample was fabricated and experiments of contact position and force detection were performed on the sample. The results showed that the principal function of the tactile sensor was feasible. The sensor sample exhibited good consistency and linearity. The tactile sensor has only five lead wires and can provide the information support necessary for safe human—computer interactions.

In this paper, aiming at the problem of control and obstacle avoidance in quadrotor formation when mathematical modeling is not accurate, the artificial potential field method with virtual force is used to plan the obstacle avoidance path of quadrotor formation to solve the problem that the artificial potential field method may fall into local optimal. The adaptive predefined-time sliding mode control algorithm based on RBF neural networks enables the quadrotor formation to track the planned trajectory in a predetermined time and also adaptively estimates the unknown interference in the mathematical model of the quadrotor to improve the control performance. Through theoretical derivation and simulation experiments, this study verified that the proposed algorithm can make the planned trajectory of the quadrotor formation avoid obstacles and make the error between the true trajectory and the planned trajectory converge within a predetermined time under the premise of adaptive estimation of unknown interference in the quadrotor model.