Xiangyu Yin’s Post

Our December issue of #Matter is now available! On the cover: The image presents an optimization landscape, with each local peak representing a unique microstructure configuration. This visual concept aligns with Steve Kench et al.'s recent paper, published in this issue, which outlines the application of generative AI in mapping processing parameters to electrode microstructures, ultimately enhancing cell performance. The landscape is depicted at sunrise, symbolizing the dawning of a new era in materials design. Image courtesy of the authors. https://lnkd.in/dTiDbJJ

🪄 How can we discover features on multiple length scales? In this publication Aditya Raghavan explores the combination of VAE and SIFT approaches for exploring how does the complexity of microstructure evolves as a function of length scale at which is is defined. Important for building materials descriptors, and important as a first step for building automated experiment workflows based on deep kernel learning. https://lnkd.in/er_qJfbZ

#AWSSageMaker & #ComputationalChemistry Manganese (Mn) compounds have emerged as potential alternatives to gadolinium-based contrast agents for Magnetic Resonance Imaging (MRI) due to their favorable relaxation properties and lower toxicity. The computational study developed at the MoMA Lab computational laboratory allowed us to model the design of a new compound with exceptional relaxivity properties using a multi-technique approach (AI Approach – DFT – Molecular Dynamics), optimizing the molecular design of a compound we had previously studied (https://lnkd.in/d-i4NnR4). We used #AWSSageMaker to evaluate the molecular design, typically by creating a SageMaker instance and installing all required libraries and frameworks, such as #RDKit for computational chemistry or #PyTorch for machine learning. Molecular data were provided in the form of SMILES strings and uploaded to an #S3bucket for easy access within #SageMaker. We used #PyTorch to train the model using molecular data. #PyTorch was integrated with #LAMMPS by building a neural network capable of predicting certain properties of the molecular system (in our case, the residence time of water coordinated around the metal center). The result was extraordinary: iterative design optimization leds to a 560% improvement in the residence time of the coordinated water. A real record!

Hello, connections! I’m extremely excited to share my second publication, "High Throughput Screening of New Piezoelectric Materials Using Graph Machine Learning and Knowledge Graphs," now published in the Computational Materials Science journal!  Computational methods, such as the Density Functional Theory (DFT), have long been a reliable tool for predicting material properties. However, their use in high-throughput screening has been limited due to computational costs. In this paper, we present a graph-based machine learning (ML) framework that overcomes these limitations, offering a more efficient approach to material selection and property prediction. Our framework, which includes a knowledge graph (KG) approach, and a graph neural network (GNN) based model, significantly reduces the search space by filtering materials from the Crystallography Open Database (COD) using KGs. Overall, we developed a pipeline for the discovery of novel piezoelectric materials to aid experimental material scientists. I want to express my heartfelt gratitude to my co-author and mentor Archit Anand for allowing me to work with him and embarking my interest in machine learning and computational materials science —your guidance and insights were invaluable.  Check out the publication https://lnkd.in/gSTUBZeu and reach out if you're interested in discussing applications or collaborating on related topics! #Publication #ComputationalMaterialsScience #MachineLearning #PiezoelectricMaterials #GraphLearning #Research #KnowledgeGraph

Continuing talking about Kolmogorov-Arnold Networks (KANs), this architecture has shown promise in several applications 𝐃𝐚𝐭𝐚 𝐅𝐢𝐭𝐭𝐢𝐧𝐠 𝐚𝐧𝐝 𝐏𝐃𝐄 𝐒𝐨𝐥𝐯𝐢𝐧𝐠 Smaller KAN models can achieve comparable or better accuracy than larger Multi-Layer Perceptron (MLP) models in tasks like data fitting and solving partial differential equations (PDEs). KANs exhibit faster neural scaling laws compared to MLPs. 𝐓𝐢𝐦𝐞 𝐒𝐞𝐫𝐢𝐞𝐬 𝐅𝐨𝐫𝐞𝐜𝐚𝐬𝐭𝐢𝐧𝐠 KANs have been applied to time series forecasting, leveraging their adaptive activation functions for enhanced predictive modeling. In a real-world satellite traffic forecasting task, KANs outperformed conventional MLPs, providing more accurate results with fewer learnable parameters.  𝐈𝐧𝐭𝐞𝐫𝐩𝐫𝐞𝐭𝐚𝐛𝐢𝐥𝐢𝐭𝐲 𝐚𝐧𝐝 𝐒𝐜𝐢𝐞𝐧𝐭𝐢𝐟𝐢𝐜 𝐃𝐢𝐬𝐜𝐨𝐯𝐞𝐫𝐲 The spline-based activation functions in KANs can be easily visualized and interpreted. KANs are useful "collaborators" that help scientists (re)discover mathematical and physical laws in examples from mathematics and physics.  𝐏𝐨𝐰𝐞𝐫 𝐒𝐲𝐬𝐭𝐞𝐦 𝐄𝐧𝐠𝐢𝐧𝐞𝐞𝐫𝐢𝐧𝐠 KANs have been applied in various applications of power system engineering, though specific details were not provided in the given search results. The flexibility, accuracy, and interpretability of KANs make them promising alternatives to traditional MLPs in a variety of domains, with potential for further development and application. For more information, you can read the original paper here: https://lnkd.in/eBCCk4JK #KAN #MachineLearning #ArtificialIntelligence #NeuralNetworks #MLP

"UCLA researchers have conducted an in-depth analysis of nonlinear information encoding strategies for diffractive optical processors, offering new insights into their performance and utility. Their study, published in Light: Science & Applications, compared simpler-to-implement nonlinear encoding strategies that involve phase encoding with the performance of data repetition-based nonlinear information encoding methods, shedding light on their advantages and limitations in the optical processing of visual information. Diffractive optical processors, built using linear materials, perform computational tasks through the manipulation of light using structured surfaces. Nonlinear encoding of optical information can enhance these processors' performance, enabling them to better handle complex tasks such as image classification, quantitative phase imaging, and encryption." #opticalprocessor #photonics

“However, such information, acquired in the form of #X-ray fluorescence spectra, is stored in complex volumes of analytical data, the examination of which represents a significant challenge in many cases,” explains Francesco Paolo Romano of Cnr-Ispc, one of the authors of the research. “The study presents a deep learning #algorithm trained on a vast synthetic dataset, composed of over 500,000 XRF spectra of pigments and #pictorial mixtures generated through #MonteCarlosimulations, a computational method used to estimate real physical quantities based on randomly generated numbers. This analytical approach based on #ArtificialIntelligence allows us to analyze accurately and precisely the millions of XRF spectra that typically make up an #MA-XRF measurement, overcoming the known limitations of conventional analysis techniques”. Artificial Intelligence sheds light on Raphael’s painting technique 👇 https://lnkd.in/dEdC_YH8 Image: RGB composition of the distribution of mercury (red), copper (blue) and iron (green) obtained from the XRF scan of a detail of the face of the fragment representing the Eternal Father by #Raphael.

Thrilled to share that my paper, "An Approach to Predict Optimal Configurations for LDA-Based Topic Modeling," has been published in the conference proceedings of Engineering Applications of Neural Networks (EANN 2024). This work focuses on optimizing hyperparameters in LDA topic modeling for better performance. Special thanks to Professor Doina Logofătu for her invaluable guidance and support. #Research #Publication #EAAN #EAAAI

✨ Thrilled to announce the publication of our latest research paper in Measurement, Elsevier! 📜 Title: "Student’s t-uniform mixture-based robust sparse coding model for sign language recognition from thermal images" Read the paper here : https://lnkd.in/e6EAVk8J This work, co-authored with Saibal Ghosh under the guidance of Dr. Amitava Chaterjee, introduces a robust gesture detection system using far-infrared thermal imaging for human-robot interaction (HRI). 🚀 Highlights of our research: We propose a far-infrared thermal imaging based robust gesture detection system in human-robot interaction (HRI) under challenging environmental conditions. We introduce a novel Student’s t-uniform mixture error distribution model to handle coding error residuals effectively and evaluate it with the standard gestures from American manual alphabet library. This distribution model's strong peak at zero and elongated tails enhance robustness against higher residuals. Case studies with normal and pixel-degraded thermal sign images under low light, occlusion, and noise demonstrate the system's resilience and effectiveness in such scenarios. 📚 Grateful for the opportunity to contribute to such a transformative field and for the collaboration that made it possible! #Research #SignLanguageRecognition #ThermalImaging #HumanRobotInteraction #MachineLearning #HumanComputerInteraction

🚀 Our paper about Physics-Informed version of Kolmogorov–Arnold Networks has been published in Computer Methods in Applied Mechanics and Engineering (CMAME)! Title: "A Physics-Informed Deep Learning Framework for Solving Forward and Inverse Problems Based on Kolmogorov–Arnold Networks (KANs)" 🔍 What’s it about? AI for partial differential equations (PDEs) is rapidly evolving, PINNs leading the change. In this work, we introduce the Kolmogorov–Arnold-Informed Neural Network (KINN), which leverages the interpretability and efficiency of KANs to improve PDE solutions in forward and inverse problem settings. We systematically explore KANs versus traditional MLP-based PINNs across various PDE scenarios, including: ✅ Multi-scale problems ✅ Singularities ✅ Stress concentrations ✅ Nonlinear hyperelasticity ✅ Heterogeneous materials ✅ Complex geometries, such as attached figure 💡 Key insights: Better accuracy and convergence: KINN significantly outperforms traditional MLPs for most PDE applications in computational solid mechanics. Efficiency with fewer parameters: KINNs achieve higher accuracy with reduced computational overhead, highlighting their potential for scalable AI-based PDE solutions. Limitations: While KINN excels across most cases, complex geometry problems remain an area for further exploration. Thanks to Yizheng Wang due to this great work. 📄 Check out the full paper here: https://lnkd.in/e3t589cE