Publications

The Computational Crystallography Toolbox (CCTBX) is open-source software that allows for processing of crystallographic data, including from serial femtosecond crystallography (SFX), for macromolecular structure determination. We describe the potential impact of using self-supervised deep learning to correct the scientific model in CCTBX and provide uncertainty quantification. We describe open questions in algorithm development to help spur advances through dialog between crystallographers and machine learning researchers.

Vidya Ganapati, Daniel Tchon, Aaron S. Brewster, Nicholas K. Sauter

ICML 2023 SynS & ML Workshop

| Link | PDF | Code |

This work creates a novel method for data-driven reconstruction in computational imaging, when no ground truth training dataset is available. Light-emitting diode (LED) array microscopy, a modality that allows visualization of transparent objects in two and three dimensions with high resolution and field-of-view, is used as an illustrative example. Future directions are outlined for making this method applicable to large-scale scientific problems.

Andrew Olsen, Yolanda Hu, Vidya Ganapati

NeurIPS 2022 AI for Science Workshop

| Link | PDF | Code | Data |

Computed tomography allows for the elucidation of 3-dimensional internal structure by the attenuation of x-rays through an object at different rotations relative to the beam. However, the x-ray beam can damage the sample, necessitating the development of sparse x-ray tomography that only requires measurements at a low number of rotation angles. This work develops and validates a self-supervised probabilistic deep learning technique for object reconstruction in this sparse regime.

Rey Mendoza, Minh Nguyen, Judith Weng Zhu, Vincent Dumont, Talita Perciano, Juliane Mueller, Vidya Ganapati

NeurIPS 2022 Machine Learning and the Physical Sciences Workshop

| Link | PDF | Code |

Human calf bioimpedance can be a useful for diagnostics and monitoring in conditions such as congestive heart failure. We prototype a fabric sock to mimic a printed circuit board and electronically connect to the calf; the sock consists of layers of conductive and insulating fabric connected by metallic snaps that act as vias.

Maggie Delano, Vidya Ganapati, Rezhwan Kamal, Bryan Le, Jackie Le, Rey Mendoza

Frontiers in Electronics (2022)

| Link | PDF |

What are the best hyperparameters (e.g. batch size, training rate, network depth) for a neural network? We would like to optimize these hyperparameters, but obtaining the derivatives for a gradient-based search is difficult. Thus, we estimate surrogate models of the underlying function relating the hyperparameters to figure of merit. Using forward function calls, we improve this surrogate function until our compute budget runs out. We validate HYPPO for different use cases, including sparse x-ray computed tomography.

Vincent Dumont, Casey Garner, Anuradha Trivedi, Chelsea Jones, Vidya Ganapati, Juliane Mueller, Talita Perciano, Mariam Kiran, Marc Day

2021 IEEE/ACM Workshop on Machine Learning in High Performance Computing Environments

| Link | PDF | Code |

See code documentation here.

Thermophotovoltaics (TPV) convert radiative heat to electricity with photovoltaics cells. This is a generalization of solar photovoltaics, which convert the radiation of the sun to electricity. TPV has the potential for high-efficiency electrical power generation from the combustion of fuel. TPV can also be used for converting waste heat or stored thermal energy to electricity. We achieved a record 28.8% efficiency of heat to electricity conversion by using photovoltaic cells with high back reflectivity.

Zunaid Omair, Gregg Scranton, Luis M. Pazos-Outón, T. Patrick Xiao, Myles A. Steiner, Vidya Ganapati, Per F. Peterson, John Holzrichter, Harry Atwater, Eli Yablonovitch

Proceedings of the National Academy of Sciences (2019)

| Link | PDF |

An experimental validation of the methods proposed and tested in simulation in Robey and Ganapati.

Yi Fei Cheng, Megan Strachan, Zachary Weiss, Moniher Deb, Dawn Carone, Vidya Ganapati

Optics Express (2019)

| Link | PDF |

Computational imaging concerns the joint design of physical hardware with image post-processing algorithms. The data recorded by the image sensor may be garbled, but as long as it contains the necessary information, the final image can be reconstructed with image post-processing. The advantages of computational imaging are reduced hardware costs and the possible recovery of quantities difficult to image directly, such as depth and the phase of light. We optimize the end-to-end computational imaging pipeline with deep learning, in order to surpass existing imaging tradeoffs among speed, resolution, and field-of-view.

Alexander Robey, Vidya Ganapati

Optics Express (2018)

| Link | PDF |

We proposed low-index intermediate “mirrors” to increase the efficiency of multi-bandgap solar cells. This theoretical result was validated with a 38.8% efficiency 4-bandgap solar cell developed at the National Renewable Energy Laboratory, setting the efficiency record for flat-plate solar cells.

Vidya Ganapati, Chi-Sing Ho, Eli Yablonovitch

IEEE Journal of Photovoltaics (2014)

| Link | PDF |

Solar cells are routinely textured in order to increase light absorption and efficiency. The benefits of texturing are generally evaluated using ray optics approximations. These approximations fail when the solar cell thickness is on the order of the wavelength of light. We use inverse design and shape calculus to optimize surface textures for subwavelength-thick solar cells. We show that with optimized designs we can approach ray-optics limits for these thin cells.

Vidya Ganapati, Owen D. Miller, Eli Yablonovitch

IEEE Journal of Photovoltaics (2014)

| Link | PDF |