Machine Learning Materials Science Postdoctoral Researcher

Berkeley Lab: building the computational mind for the 'swiss army knife' of microscopes. August 18, 2026 Press play to listen to this content Aug. 18, 2026 - Biology doesn't happen at one scale. Molecular interactions unfold in milliseconds and nanometers, while disease-associated change such as in Alzheimer's spreads across millimeters of brain tissue. Understanding complex biological systems requires scientists to watch both - ideally at the same time and with the same sample. Historically, this has meant shuttling samples between specialized instruments, often damaging biological context and slowing results. There's also a common crux across microscopes: the closer you look at living tissue, the more the image blurs, and the more detail you capture, the more overwhelming the resulting data becomes. Five imaging modes provide complementary views of the same dividing human retinal pigment epithelial cell (hTERT-RPE1). From top left, clockwise: Lattice Light-Sheet Structured Illumination Microscopy (LLS-SIM), widefield, 3D-SIM, oblique illumination, lattice light sheet. This shows how different microscopy techniques compare when imaging the same cell, giving them a more complete picture of what's happening inside. (Credit: Fu, Liu, Milkie, Ruan et al., Nature Methods, 2026) A new instrument aims to address these problems - and has revealed a third, arguably harder challenge that Berkeley Lab is uniquely positioned to address. Researchers at Lawrence Berkeley National Laboratory (Berkeley Lab) and collaborating institutions have developed the Multimodal Optical Scope with Adaptive Imaging Correction (MOSAIC) - a reconfigurable microscope that consolidates more than ten imaging techniques into one compact instrument. It processes its massive datasets using computational tools developed at Berkeley Lab, funded by a Laboratory Directed Research and Development (LDRD) award and supported by the Perlmutter supercomputer at the National Energy Research Scientific Computing Center (NERSC). Featured on the cover of Nature Methods, MOSAIC allows scientists to track biological processes across scales and compare imaging methods on the same sample. It generates data at a pace that is pushing the boundaries of what biology can discover. "MOSAIC can generate up to four terabytes of data per hour - far beyond what conventional processing workflows or human inspection can handle," said Srigokul "Gokul" Upadhyayula, a faculty scientist in the Molecular Biophysics and Integrated Bioimaging Division and co-corresponding author. "The microscope is only as useful as our ability to process those data and extract biological meaning from them. Berkeley Lab's expertise in high-performance computing and large-scale data analysis is essential to closing that gap." How MOSAIC Works MOSAIC grew from the adaptive-optical lattice light-sheet microscope reported by Nobel laureate Eric Betzig, Upadhyayula, and their colleagues in 2018. That earlier system delivered exceptional performance but was so large it occupied a 10-foot by 4-foot optical table. As demand from the broader research community grew, the team designed the MOSAIC to retain and expand those capabilities while reducing the instrument's footprint. "MOSAIC has been built over a dozen times in different places with over 50 research licenses already shared. We also created comprehensive documentation on how to build this instrument - think an IKEA-style instruction set geared towards a scientist who has no deep optical expertise but is willing to learn." MOSAIC's main innovation is that it can be quickly reconfigured in two to five seconds to switch between a dozen distinct imaging modes. The team designed a smart modular system where the same lasers, mirrors, cameras, and computational hardware serve multiple imaging functions through a custom optical switching system. Critically, every one of those modes is enhanced with adaptive optics: a technology borrowed from astronomers who developed it to sharpen images of distant stars blurred by Earth's atmosphere. This corrects blurring caused by aberrations in the living tissue itself. "Sample-induced aberrations distort and redirect light, reducing both signal and resolution," said Upadhyayula. "Adaptive optics measures those distortions and corrects them. It is like turning on the windshield wipers while driving in the rain: the information is present all along, just obscured." MOSAIC also relies on fluorescent molecules that allow biologists to mark specific cellular structures and molecular activities in living cells, fast and gentle light-sheet imaging that captures cellular dynamics with minimal stress or damage, and high-speed data transfer infrastructure capable of moving and processing massive imaging datasets. What Becomes Visible When You Clear the Windshield? MOSAIC's ability to image with minimal invasiveness at large scales over long durations has already enabled several experiments: tracking single molecules in living cells, observing organelle dynamics in developing zebrafish embryos, mapping neuronal architecture in expanded human brain tissue from a person with Alzheimer's, and imaging neural activity in live mouse brains. In that last application, adaptive optics correction revealed roughly 2.5 times more detectable neural calcium events than imaging without it - suggesting conventional microscopy has been quietly undercounting brain activity. MOSAIC is also the instrument that powered a related study on Volumetric Imaging via Photochemical Sectioning (VIPS), published in Science in 2025. That project used MOSAIC and the computational tools developed at Berkeley Lab to image two complete adult mouse olfactory bulbs at nanoscale resolution, generating roughly a petabyte of data in approximately two weeks. Analyzing it took two years: an illustration of the gap between what these instruments can see and what researchers can currently process. "The bottleneck is no longer our ability to acquire the data," said Upadhyayula. "These microscopes can generate massive datasets at staggering rates. The key bottleneck is turning dense five-dimensional observations into biological understanding." Berkeley Lab's contribution helps to target this gap. Round-the-Clock Data Collection for Biological AI Supported by an LDRD award, Eric Betzig and Upadhyayula's group developed PetaKit5D, an open-source software toolkit that can handle MOSAIC's terabyte-per-hour output in real time and cuts processing costs by more than an order of magnitude compared to previous approaches. The team also secured computing allocations on the Perlmutter supercomputer at NERSC to process and visualize portions of the largest datasets. But processing data efficiently is only half the equation. The other half is generating enough of it - consistently, at scale, and of sufficient quality - to train the kind of AI model that could one day make sense of it all. At UC Berkeley, two MOSAIC instruments now run around the clock, capturing the five-dimensional data - three spatial dimensions, time, and molecular identity - that will be needed to train a new state-of-the-art AI model. The data flowing from those instruments already represents a fundamental shift in how biology can be practiced. For the first time, researchers can watch in vivo biochemistry unfold inside cells living within their native tissues, inside a living organism. What comes next, he believes, could be transformative: a vision language model that reasons natively over biology, connecting what it sees with molecular identity, experimental context, and prior biological knowledge to determine which observations matter and which experiments should come next. "Connected to automated microscopes, sample handling, and perturbation systems, that capability could provide the foundation for self-driving biological laboratories - and fundamentally change the rate at which we can make discoveries," said Upadhyayula. Deep Origin this month announced that its AI drug discovery framework delivered nearly a 31%... AI models are getting better at a rapid pace. They are now able to reason,... The tech industry is fixated on one main metric, the raw number of GPUs accumulated... Scientists running large simulations can end up with terabytes of data that then has to... Jensen Huang believes NVIDIA's chips are becoming much more than expensive pieces of hardware. As... As hybrid cognition deepens, the boundary between biological and artificial embodiment will begin to blur...

Predicting microelectronics performance with physics-informed artificial intelligence. Aug 06, 2026 A new AI framework will link atomic-scale defects to device performance, helping researchers detect failures earlier and design more reliable, efficient electronics. (Nanowerk News) How a microelectronic device performs depends on what it is made of, how well its materials and interfaces are built and how electricity, heat and tiny defects change over time. These devices power much of modern life - from smartphones and laptops to secure communications and artificial intelligence (AI) hardware. As next-generation devices become smaller, faster and more tightly packed, their performance is affected more and more by tiny flaws in materials and interfaces. These defects can lead to overheating, electrical leakage, unreliable switching and, in the end, shorter device lifetimes. But defects are not always harmful. They can also influence electrical and thermal behavior in useful ways, depending on how they are distributed and how they evolve over time. Understanding both the harmful and beneficial effects of defects is essential for designing better microelectronics. To address the challenges defects pose, researchers at the U.S. Department of Energy's (DOE) Argonne National Laboratory, Lawrence Berkeley National Laboratory (Berkeley Lab), Oak Ridge National Laboratory (ORNL) and Northwestern University plan to develop the Materials Discovery Cloud. The project will create a physics-informed AI framework that learns how material composition, structure and operating conditions influence defect evolution and key functional properties such as electrostatic potential, current density and temperature. From protein folding to device function. AlphaFold is an AI system developed by Google DeepMind that predicts a protein's 3D structure based on its amino acid sequence. It transformed biology by making structure prediction much faster and more accurate, helping scientists better understand how proteins work. The new microelectronics effort follows a similar idea but applies it to a very different problem. Instead of predicting protein structure from sequence, the team aims to predict how networks of defects form, change and affect how a device works. "In biology, AlphaFold learned to connect sequence to structure," said Subramanian Sankaranarayanan, Argonne scientist and lead principal investigator on the project, as well as a professor at the University of Illinois Chicago. "We want to connect defect distributions in materials and interfaces to the electrical and thermal properties that matter for microelectronics. Such a framework remains elusive." Why defects are so hard to understand. Defects are tiny irregularities in a material's structure. They can include missing atoms, dislocations, voids or chemical disorder. Some defects hurt performance. Others can help enable useful behavior. The challenge is knowing which defects matter, when they matter and how they change under real operating conditions. These attributes are difficult to determine because no single instrument can capture the whole picture. Some tools, such as electron microscopes, can directly image features at very small scales. X-ray methods can reveal strain, buried structures and defect motion. Other techniques measure chemistry, electrical behavior and heat flow. Each method shows one part of the story, but not the whole system. A useful way to think about it is like trying to understand the day's weather from the temperature alone. Temperature tells you something important, but you also need to factor in wind, clouds, precipitation and humidity to see the full picture. The same is true for microelectronics. Bringing many tools into one framework. To close that gap, the team is combining many types of data from DOE Office of Science user facilities and advanced computing systems, drawing information from different tools and scales into one unified platform. At Argonne, those include the Advanced Photon Source and the Center for Nanoscale Materials for X-ray and microscopy measurements, as well as the Argonne Leadership Computing Facility for large-scale computing. Partner capabilities include the Advanced Light Source, the Molecular Foundry and the National Energy Research Scientific Computing Center at Berkeley Lab, and the Center for Nanophase Materials Sciences at ORNL. The Materials Discovery Cloud will help researchers gather experimental data, run simulations and generate synthetic data that mimic experiments. The synthetic data is especially important because complete experimental data sets for material samples are still complicated and time-consuming to collect. Simulations can fill in missing pieces and help train the AI framework. Another part of the effort focuses on autonomous discovery, which uses AI, machine learning and robotics to help researchers decide which measurements to run next and collect new data more quickly. That can help address one of the project's key challenges: generating enough high-quality experimental data to build and refine AI models. In this work, researchers are developing an AI-guided platform for microelectronics materials that can synthesize samples and carry out multiple kinds of characterization in a more integrated, high-throughput workflow. By reducing the need to move samples among facilities for separate measurements, the approach could speed data collection and help fill important gaps in the multimodal datasets used to train the framework. "What makes the Materials Discovery Cloud powerful is that we can bring together experiments, simulations and AI in one workflow," Sankaranarayanan explained. "That gives us a way to learn from limited data today while building a framework that can grow more capable as new data comes in." AI guided by physics. This system is not designed to be a black box that gives answers without explanation. Instead, it will be built around well-established laws of physics. This helps ensure the AI's predictions are grounded in how materials and devices actually behave. That certainty is important because the researchers want the system to do more than spot patterns in data. They want it to help reveal why certain defects lead to specific changes in performance, reliability or lifetime. To do that, the framework will combine several kinds of AI tools that can bring together many types of data, learn from both experiments and simulations and identify which new measurements would be most useful next. In the end, the goal is to connect tiny, atomic-scale features in a material to the larger electrical, thermal and mechanical behavior of a real device. What success could look like. If successful, the Materials Discovery Cloud could change how scientists design and test new materials and devices. Instead of waiting through long rounds of experiments, researchers may be able to get useful answers from a smaller set of early measurements. That could help them spot problems sooner, avoid spending time on weak candidates and focus more quickly on the most promising designs. Over time, the framework could also support inverse design. Rather than starting with a material and seeing how it performs, they could start with a goal - such as better heat management or a longer-lasting device - and ask what kind of material structure or defect pattern would be needed to achieve it. By bringing many different measurements into one predictive system, the effort aims to help scientists better understand, control and design the materials behind the next generation of microelectronics.