A neural network learns to sculpt materials at the nanoscale, replacing the slow trial-and-error of focused ion beam milling with a system that predicts the exact settings needed to carve a desired structure. Current nanofabrication relies on guesswork: engineers run dozens of test patterns to calibrate the ion beam for each new material or geometry, wasting time and material. The underlying physics of how ions carve into a surface is too complex to model from first principles. This project replaces that guesswork with a deep-learning model that learns the relationship between beam settings and the resulting shape directly from data, for any metal, semiconductor, or dielectric. If successful, the technique could accelerate prototyping of metasurface optics—ultra-thin lenses for cameras and sensors—and optomechanical devices that control light with vibrations. It would also speed up sample preparation for electron microscopy, a bottleneck in materials science and semiconductor failure analysis. The method is designed to transfer to any focused ion beam platform and to other direct-write fabrication tools, making nanoscale manufacturing faster, more reproducible, and less wasteful across UK research labs and high-tech industries.
View original technical description
We seek to improve the efficacy, efficiency and reproducibility of focused ion beam (FIB) nanofabrication processes for advanced photonic materials and devices, and in turn their optical performance and energy efficiency, through the application of deep learning. Our methods will enable rapid optimization of nanostructures for optical function informed by real process and material characteristics, rather than analytical/numerical approximations or wasteful trial-and-error. Deep learning offers a novel, systematic, data-driven approach to modelling the FIB milling process that is applicable to arbitrary nano/microstructural geometries in any target medium, whereby milling outcomes can be accurately predicted without the need for knowledge or understanding of (often unavailable/inaccessible) material parameters, the fundamental physics of ion-atom interactions, or the mathematical description of time-dependent 3D structural geometry. We will show how neural networks can: rapidly accrue understanding of the complex relationships among numerous sample and system parameters that affect process outcomes for a variety of metal, semiconductor and dielectric target materials commonly used in nanophotonic devices, to expedite or negate the need for conventional 'dose testing'; be configured to solve challenging inverse problems, answering the question "what input design will generate a desired output structure?"; optimize (rather than just simulate) milling processes, to answer the question "what input design and process parameters will best deliver a structure with prescribed optical and/or nanomechanical properties?". Alongside these trained functionalities we expect emergent capabilities, such as the ability to make accurate predictions for unseen target media, and inherent compensation for systematic artefacts. The project aligns to the EPSRC strategic priority on "Frontiers in Engineering and Technology": it will leverage novel capabilities and ideas in deep learning and nanophotonics to facilitate "breakthroughs in ... tools and techniques enabling researchers and businesses to make, measure and model more efficiently and effectively," and to "accelerate design to manufacture of the new materials needed for a more resilient, sustainable UK". FIB milling is a key enabling technology in many areas of fundamental and applied research, and high-tech (electronics and photonics) industrial applications of importance to the UK's position at the forefront of physical and bio sciences research and advanced technology development - for micro/nanoscale rapid prototyping, materials/device characterisation and cross-sectional/tomographic imaging (including in support of other high-throughput, e.g. lithographic, nanofabrication processes), and transmission electron microscopy sample preparation. Our initial focus will be on applications in nanophotonics (e.g. to the optimized fabrication of metasurface optics and energy-efficient optomechanical time crystals), but the techniques developed will be transferrable to beneficiaries in other domains of engineering, physical science and technology, deployable on any FIB platform, and indeed adaptable to other direct-write micro/nanofabrication processes.
Plain English summaries and category classifications on this site are generated by AI and may not perfectly reflect the original research.
Is something wrong? Let us know