Alfonso R. Reyes’ Post

The integration of #AI into physics research represents a monumental shift, accelerating scientific progress across various disciplines. AI’s ability to process vast amounts of #data and simulate complex systems has revolutionized how physicists approach #dataanalysis, experimental design, and discovery. From enhancing data analysis capabilities to enabling predictive modeling and driving discoveries, AI offers unprecedented opportunities to push the boundaries of human knowledge in fundamental physics and beyond. https://lnkd.in/diVa4jHD #worlddatasummit #machinelearning #dataanalytics #dataliteracy #dataarchitecture #dataquality #bigdata #datainnovation #dataethics #businessintelligence

"The hype surrounding machine learning, a form of artificial intelligence, can make it seem like it is only a matter of time before such techniques are used to solve all scientific problems. While impressive claims are often made, those claims do not always hold up under scrutiny. Machine learning may be useful for solving some problems but falls short for others. In a new paper in Nature Machine Intelligence, researchers at the U.S. Department of Energy's Princeton Plasma Physics Laboratory (PPPL) and Princeton University performed a systematic review of research comparing machine learning to traditional methods for solving fluid-related partial differential equations (PDEs). Such equations are important in many scientific fields, including the plasma research that supports the development of fusion power for the electricity grid." #ai

Researchers at KAIST have developed DeepSCF, an AI-based model that speeds up complex quantum mechanical calculations by eliminating the repetitive, computationally intense self-consistent field (SCF) processes required in traditional density functional theory (DFT). Using a neural network trained on 3D chemical bonding data, the model accurately predicts quantum properties across large systems with less computational effort. This breakthrough allows faster, scalable simulations in materials science and drug design, showing promise for enhanced AI applications in quantum mechanics. For more details, please continue reading the full article under the following link: https://lnkd.in/er7DdPUx -------------------------------------------------------- Please consult also the Quantum Server Marketplace platform for the outsourcing of computational science R&D projects to external expert consultants through remote collaborations: https://lnkd.in/eRmYbj4x #materials #materialsscience #materialsengineering #computationalchemistry #modelling #chemistry #researchanddevelopment #research #MaterialsSquare #ComputationalChemistry #Tutorial #DFT #simulationsoftware #simulation

Researchers at KAIST have developed DeepSCF, an AI-based model that speeds up complex quantum mechanical calculations by eliminating the repetitive, computationally intense self-consistent field (SCF) processes required in traditional density functional theory (DFT). Using a neural network trained on 3D chemical bonding data, the model accurately predicts quantum properties across large systems with less computational effort. This breakthrough allows faster, scalable simulations in materials science and drug design, showing promise for enhanced AI applications in quantum mechanics. For more details, please continue reading the full article under the following link: https://lnkd.in/er7DdPUx -------------------------------------------------------- Please consult also the Quantum Server Marketplace platform for the outsourcing of computational science R&D projects to external expert consultants through remote collaborations: https://lnkd.in/eRmYbj4x #materials #materialsscience #materialsengineering #computationalchemistry #modelling #chemistry #researchanddevelopment #research #MaterialsSquare #ComputationalChemistry #Tutorial #DFT #simulationsoftware #simulation

Researchers at KAIST have developed DeepSCF, an AI-based model that speeds up complex quantum mechanical calculations by eliminating the repetitive, computationally intense self-consistent field (SCF) processes required in traditional density functional theory (DFT). Using a neural network trained on 3D chemical bonding data, the model accurately predicts quantum properties across large systems with less computational effort. This breakthrough allows faster, scalable simulations in materials science and drug design, showing promise for enhanced AI applications in quantum mechanics. For more details, please continue reading the full article under the following link: https://lnkd.in/eptpjYut -------------------------------------------------------- Please consult also the Quantum Server Marketplace platform for the outsourcing of computational science R&D projects to external expert consultants through remote collaborations: https://lnkd.in/eCb9Tanv #materials #materialsscience #materialsengineering #computationalchemistry #modelling #chemistry #researchanddevelopment #research #MaterialsSquare #ComputationalChemistry #Tutorial #DFT #simulationsoftware #simulation

Catalysis in the digital age: leveraging machine learning for catalyst discovery The discovery and design of new catalysts are critical for advancing technologies in energy conversion, environmental remediation, and the chemical industry. However, traditional catalyst development, often based on trial and error, faces significant challenges due to the complexity of catalytic reactions and material properties. Recently, machine learning (ML) has emerged as a transformative tool, enabling data-driven approaches to predict key properties such as stability, activity, and selectivity. In a recent review, Bokinala Moses Abraham and colleagues explore the use of ML algorithms like neural networks, random forests, and support vector machines to revolutionize computational catalysis. By integrating experimental data with theoretical models such as density functional theory (DFT), ML models efficiently identify optimal catalysts for various reactions. Notably, these algorithms excel in feature engineering, identifying key patterns in high-dimensional datasets to predict reaction pathways and material properties. For example, gradient boosting and deep learning methods have demonstrated exceptional accuracy in predicting adsorption energies and reaction barriers, key factors in catalytic efficiency. A key highlight of this work is its focus on data-driven solutions to overcome traditional computational bottlenecks, such as the high cost and complexity of DFT calculations for large systems. The authors highlight the creation of comprehensive scientific data repositories that facilitate the training of robust ML models capable of navigating complex optimization problems in catalyst design. These databases, combined with advanced ML techniques like Gaussian process regression and genetic algorithms, enable researchers to optimize catalyst performance with reduced computational costs. Review paper (open access): https://lnkd.in/dctRVtjc #MachineLearning #Catalysis #ArtificialIntelligence #NeuralNetworks #SupportVectorMachines #GradientBoosting #CatalystDiscovery #AIforScience #EnergyConversion #Sustainability #MaterialsScience #DataScience #DFT #InnovationInChemistry

A new technique that can automatically classify phases of physical systems could help scientists investigate novel materials. The MIT team’s insight is that this probability distribution also defines a generative model upon which a classifier can be constructed. They plug the generative model into standard statistical formulas to directly construct a classifier instead of learning it from samples, as was done with discriminative approaches. #GenAI

Unsolicited advice for young scientists doing computational modeling: Remember that numerical algorithms are mathematical tools that do not understand physics. Newton’s method can converge to spurious (non-physical) solutions that satisfy your set of equations and neural networks and optimization algorithms can “bend” the laws of physics in order to fit your data or find an optimum. Do not trust solutions delivered by an algorithm blindly. A lot of effort has to be spent to validate that the solution makes physical and practical sense. In addition, it is important to understand/explain why the algorithm is selecting a particular solution. I also think that it is important to remember that certain aspects arising in real applications cannot be easily expressed mathematically (fairness, sustainability, consensus, happiness, risk, human behavior). As such, solutions delivered by an algorithm need to be scrutinized. #phdresearch #computation

Good advice on computational modeling. And the same goes for computational optimization: your model doesn't understand the real world. If you have forgotten an important constraint, then thats on you. I attended a talk once where the speaker presented a railway scheduling/management problem for the German railway system. One of the tasks was to find out how many carriages should be stored in which depot over night, to maximize various different performance criteria. (Basically, it makes sense to store carriages where the morning demand of commuting to the workplace occurs, right? That seems to be a good approach to at least minimize overall travel distance.) Now Germany isn't operating in a vacuum; there are various trains running to and from neighbouring nations. Now what happens when a German railway carriage ends up in France overnight? At the time of the talk, the French paid for "renting" the carriage. Handsomely, apparently. And the same for other nations. And so the optimization model dutifully pushed as many carriages as possible into neighbouring nations at the end of the day. Because thats how you made the most money. Not the solution you are looking for.

𝑳𝒆𝒕’𝒔 𝒔𝒕𝒂𝒓𝒕 𝒕𝒉𝒆 𝒘𝒆𝒆𝒌 𝒘𝒊𝒕𝒉 𝒂𝒏 𝒆𝒙𝒄𝒊𝒕𝒊𝒏𝒈 𝒖𝒑𝒅𝒂𝒕𝒆 𝒇𝒓𝒐𝒎 𝒐𝒖𝒓 𝒈𝒓𝒐𝒖𝒑! Our PhD student, Jesús González Sieiro, has just published his first peer-reviewed journal paper titled “𝐑𝐞𝐝𝐮𝐜𝐢𝐧𝐠 𝐒𝐩𝐚𝐭𝐢𝐚𝐥 𝐃𝐢𝐬𝐜𝐫𝐞𝐭𝐢𝐳𝐚𝐭𝐢𝐨𝐧 𝐄𝐫𝐫𝐨𝐫 𝐨𝐧 𝐂𝐨𝐚𝐫𝐬𝐞 𝐂𝐅𝐃 𝐒𝐢𝐦𝐮𝐥𝐚𝐭𝐢𝐨𝐧𝐬 𝐔𝐬𝐢𝐧𝐠 𝐚𝐧 𝐎𝐩𝐞𝐧𝐅𝐎𝐀𝐌-𝐄𝐦𝐛𝐞𝐝𝐝𝐞𝐝 𝐃𝐞𝐞𝐩 𝐋𝐞𝐚𝐫𝐧𝐢𝐧𝐠 𝐅𝐫𝐚𝐦𝐞𝐰𝐨𝐫𝐤” in the Engineering with Computers Journal by Springer Nature Group. I am particularly proud of Jesús’s work because it has achieved the following milestones: 1. 𝑅𝑒𝑑𝑢𝑐𝑒𝑑 𝑡ℎ𝑒 𝑒𝑟𝑟𝑜𝑟 𝑓𝑟𝑜𝑚 120% 𝑡𝑜 25% 𝑖𝑛 𝑣𝑒𝑙𝑜𝑐𝑖𝑡𝑦 for simulations within the training distribution, using an 8x coarser mesh compared to the traditional solver. 2. Developed 𝑎 𝑓𝑢𝑙𝑙𝑦 𝑒𝑚𝑏𝑒𝑑𝑑𝑒𝑑 𝐷𝐿/𝑂𝑝𝑒𝑛𝐹𝑂𝐴𝑀 𝑓𝑟𝑎𝑚𝑒𝑤𝑜𝑟𝑘. 3. Made 𝑂𝑝𝑒𝑛𝐹𝑂𝐴𝑀 𝑠𝑝𝑎𝑡𝑖𝑎𝑙𝑙𝑦 𝑑𝑖𝑓𝑓𝑒𝑟𝑒𝑛𝑡𝑖𝑎𝑏𝑙𝑒. 4. Applied 𝑑𝑒𝑒𝑝 𝑛𝑒𝑢𝑟𝑎𝑙 𝑛𝑒𝑡𝑤𝑜𝑟𝑘𝑠 𝑡𝑜 𝑎 𝑑𝑒𝑠𝑖𝑔𝑛 𝑝𝑟𝑜𝑏𝑙𝑒𝑚 𝑓𝑜𝑟 𝑒𝑛𝑔𝑖𝑛𝑒𝑒𝑟𝑠, beyond just classification and operation purposes. Jesús’s approach sets a new paradigm in #CFD. Traditionally, the ML component has been decoupled from the CFD solver, potentially leading to unsatisfied constitutive equations. However, Jesús’s method ensures fulfillment at each approximation by leveraging the local features of the physics. Coauthored with David Pardo, Victor Calo, and M. Towara, Jesús demonstrated remarkable persistence in achieving these results. While promising, these findings still require further refinement to fully harness the method’s potential. This makes Jesús an exceptional PhD candidate. Most importantly, Jesús already has ideas on how to apply this method to more realistic conditions, enhancing its utility for engineering purposes and improving accuracy. You can read and download the paper follwoing the link. Enjoy your reading! https://lnkd.in/dMZxhWTD