Estimating sample parameters from diffraction patterns is one of those inverse problems where expensive simulations limit throughput. Nearby on the archive are papers on shadow tomography, assembling mosaics from rare flashes, determining molecular shapes in a microscope, and accelerating molecular calculations with machine learning. All of them seek ways to squeeze more information from a limited number of measurements or simulations. This work follows the same vein but bets on exploiting the known structure of the objective function rather than training a neural network.
SBOCF builds a surrogate model not for the final error but for intermediate quantities. First, the diffraction image is split into 9 square patches, and for each patch the sum of pixel intensities is computed. Then two correction terms are introduced: a multiplicative one that scales the error across patches, and an additive one that captures residual discrepancy. These 11 quantities are modeled with independent Gaussian processes. Bayesian optimization selects the next parameter set by maximizing expected improvement and runs a simulation. The budget is capped at 50 simulations.
On a synthetic SrTiO3 sample with a thickness of 380 Å and a tilt of 1.5 mrad along both axes, SBOCF reduced the median final error by a factor of 290 compared to standard Bayesian optimization with expected improvement. For a thin sample of 100 Å, the gain was 47 times; for a 200 Å sample, 37 times. Thickness and tilt estimates were more accurate: thickness error reduced by up to 24 times, tilt error by up to 35 times. On experimental data, the method gave estimates consistent with previously published ones, without prior neural network training. The time to select the next point was about 0.2 minutes per iteration.
SBOCF differs from neighboring works in that it requires no training set and is not tied to a specific material or imaging conditions. Neighboring methods either sweep a parameter grid or train a neural network on a large library of simulations. Here, only the structure of the objective function and 50 simulations are used. This extends the approach's applicability to new materials and regimes without retuning.
The method can be extended to other parameters affecting the diffraction pattern: probe aberrations, atomic thermal vibration amplitudes, and thicknesses of individual layers in heterostructures. Applying SBOCF to other inverse problems with expensive simulators and high-dimensional structured outputs, such as X-ray diffraction or neutron scattering, is also promising.
The method will speed up materials characterization in transmission electron microscopy and improve the accuracy of subsequent atomic structure reconstruction. It is also useful in any task where simulation parameters must be matched to experiment under a limited computational budget.
Test SBOCF on other materials and over a wider range of thicknesses and tilts. Integrate the method into the electron microscopy workflow for automatic reconstruction tuning.
This work directly addresses the unsolved problem of fast and accurate sample parameter estimation in electron microscopy without large training sets. This is a key barrier to automating and scaling materials structural analysis.
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The paper introduces the SBOCF method for Bayesian optimization of sample parameters in electron microscopy. The method splits the diffraction image into patches and models their summaries, reducing the output dimensionality from 24,649 to 11. On synthetic and experimental data, SBOCF reduces the error compared to standard Bayesian optimization by up to 290 times and improves the subsequent reconstruction.
Where its strength is
The strength of the method lies in exploiting the structure of the objective function through intermediate quantities, which allows it to work without a training set and without being tied to a specific material.
Where it could go next
Worth trying
Why this may matter
The proposed method extends the applicability of Bayesian optimization in inverse problems with expensive simulations, where training neural networks is difficult due to limited data.
Nearby works in our archive
The archive is sparse around this work: the closest review is only 0.53, against a usual 0.64. It looks like few have drilled nearby.
The number of modeled outputs dropped from 24,649 to 11, a reduction of more than 2000 times, while the final error decreased by up to 290 times. If each simulation took an hour, a full grid search over a million points would take 114 years, but SBOCF does it in 50 hours.