BEGIN:VCALENDAR VERSION:2.0 PRODID:IEEE vTools.Events//EN CALSCALE:GREGORIAN BEGIN:VTIMEZONE TZID:America/Los_Angeles BEGIN:DAYLIGHT DTSTART:20260308T030000 TZOFFSETFROM:-0800 TZOFFSETTO:-0700 RRULE:FREQ=YEARLY;BYDAY=2SU;BYMONTH=3 TZNAME:PDT END:DAYLIGHT BEGIN:STANDARD DTSTART:20261101T010000 TZOFFSETFROM:-0700 TZOFFSETTO:-0800 RRULE:FREQ=YEARLY;BYDAY=1SU;BYMONTH=11 TZNAME:PST END:STANDARD END:VTIMEZONE BEGIN:VEVENT DTSTAMP:20261013T220243Z UID:BDF93124-77BC-47F1-BB89-9FF91E8DF826 DTSTART;TZID=America/Los_Angeles:20261029T160000 DTEND;TZID=America/Los_Angeles:20261029T173000 DESCRIPTION:Abstract:\n\nA recent report by the US Department of Energy def ines the area of scientific machine learning as\n\n“a core component of artificial intelligence (AI) and a computational technology that can be tr ained\, with\n\nscientific data\, to augment or automate human skills”\, which has “the potential to transform science and\n\nenergy research” . In this presentation\, we discuss the potential of scientific machine le arning methods to\n\nproblems in computational electromagnetics starting f rom standard electromagnetic structure analysis\n\nand multi-physics model ing in the time-domain\, employing unsupervised\, supervised and reinforce ment\n\nlearning.\n\nAn unsupervised learning strategy is based on a modif ied Physics-Informed Neural Network (PINN). PINNs\n\ndirectly integrate ph ysical laws into their loss function\, so that the training process does n ot rely on the\n\ngeneration of training ground truth data from simulation s. We introduce a physics-informed deep-\n\noperator-network (PI-DON) for Maxwell’s equations. We show applications of the PI-DON to the analysis\ ,\n\ndesign and uncertainty quantification study of finite and defect peri odic structures\, including frequency\n\nselective surfaces and metasurfac es.\n\nMoreover\, we demonstrate the impact of machine learning on the com putational modeling of radiowave\n\npropagation scenarios and reconfigurab le intelligent surfaces\, using supervised learning. We build\n\nconvoluti onal neural network models that can process the geometry of indoor environ ments\, along with\n\nphysics-inspired parameters\, to rapidly estimate re ceived signal strength (RSS) maps. Emphasis is placed\n\non the generaliza bility of these models\, which is their ability to "learn" the physics of radiowave\n\npropagation and produce accurate modeling predictions in new geometries well beyond those included in\n\ntheir training set.\n\nFinally \, the use of reinforcement learning for the optimization of reconfigurabl e intelligent surfaces will be\n\ndiscussed.\n\nSpeaker(s): Costas\n\nRoom : 204\, Bldg: David Packard Electrical Engineering Building\, 350 JANE STA NFORD WAY\, STANFORD\, California\, United States\, 94305 LOCATION:Room: 204\, Bldg: David Packard Electrical Engineering Building\, 350 JANE STANFORD WAY\, STANFORD\, California\, United States\, 94305 ORGANIZER:austinxr@stanford.edu SEQUENCE:50 SUMMARY:Stanford MTT-S/APS DL Talk: Scientific Machine Learning for Electro magnetic Field Computations URL;VALUE=URI:https://events.vtools.ieee.org/m/585032 X-ALT-DESC:Description: <br /><p class="p1"><strong>Abstract:</strong></p>\ n<p class="p1">A recent report by the US Department of Energy defines the area of scientific machine learning as</p>\n<p class="p1">&ldquo\;a core c omponent of artificial intelligence (AI) and a computational technology th at can be trained\, with</p>\n<p class="p1">scientific data\, to augment o r automate human skills&rdquo\;\, which has &ldquo\;the potential to trans form science and</p>\n<p class="p1">energy research&rdquo\;. In this prese ntation\, we discuss the potential of scientific machine learning methods to</p>\n<p class="p1">problems in computational electromagnetics starting from standard electromagnetic structure analysis</p>\n<p class="p1">and mu lti-physics modeling in the time-domain\, employing unsupervised\, supervi sed and reinforcement</p>\n<p class="p1">learning.</p>\n<p class="p1">&nbs p\;</p>\n<p class="p1">An unsupervised learning strategy is based on a mod ified Physics-Informed Neural Network (PINN). PINNs</p>\n<p class="p1">dir ectly integrate physical laws into their loss function\, so that the train ing process does not rely on the</p>\n<p class="p1">generation of training ground truth data from simulations. We introduce a physics-informed deep- </p>\n<p class="p1">operator-network (PI-DON) for Maxwell&rsquo\;s equatio ns. We show applications of the PI-DON to the analysis\,</p>\n<p class="p1 ">design and uncertainty quantification study of finite and defect periodi c structures\, including frequency</p>\n<p class="p1">selective surfaces a nd metasurfaces.</p>\n<p class="p1">Moreover\, we demonstrate the impact o f machine learning on the computational modeling of radiowave</p>\n<p clas s="p1">propagation scenarios and reconfigurable intelligent surfaces\, usi ng supervised learning. We build</p>\n<p class="p1">convolutional neural n etwork models that can process the geometry of indoor environments\, along with</p>\n<p class="p1">physics-inspired parameters\, to rapidly estimate received signal strength (RSS) maps. Emphasis is placed</p>\n<p class="p1 ">on the generalizability of these models\, which is their ability to "lea rn" the physics of radiowave</p>\n<p class="p1">propagation and produce ac curate modeling predictions in new geometries well beyond those included i n</p>\n<p class="p1">their training set.</p>\n<p class="p1">&nbsp\;</p>\n< p class="p1">Finally\, the use of reinforcement learning for the optimizat ion of reconfigurable intelligent surfaces will be</p>\n<p class="p1">disc ussed.</p> END:VEVENT END:VCALENDAR