Predicting Biophotovoltaic Current Output and Photoresponse Using Machine Learning

Abstract

Background Biophotovoltaics (BPVs) promise a low-cost, sustainable option for electricity and solar fuels production from just light as an energy source, water as an electron source, and air as an inorganic carbon source. This is achieved by harnessing the processes of photosynthesis and respiration by exoelectrogenic photosynthetic microorganisms in electrochemical cells1,2.BPVs are advantageous over microbial fuel cells in which the microbes require an external organic energy source and an oxygen-free enclosure. BPVs are also advantageous over synthetic solar PV cells in that 24-hour power production is possible courtesy of respiration in the dark, and the living organisms can self-repair damaged photoactive components. This makes them particularly suited for operation in off-grid and remote locations. Research Challenge BPV power outputs remain too low for commercially feasibility. Work on BPVs has hitherto largely been an experimental exercise, with only a handful of computational studies in the literature3. Developing computational models that can aid in interpreting experimental results, or that can be used for rapid sensitivity analyses and device optimisation, has proved challenging. This is largely due to the gaps in our knowledge of the full path electrons take from within the microorganisms, to the non-living electrodes. Approach To overcome these gaps, “deep learning” was applied to predict BPV current density and photo-response, since this approach does not require upfront definition of all the interacting electrochemical, biological and physical phenomena occurring within the system. In particular, Long Short-Term Memory (LSTM) networks were used4. Current density profiles from BPV devices operating in galvanic mode with a 33 MΩ load, under a 12h:12h on:off light cycle, were decomposed with Seasonal and Trend Decomposition using locally estimated scatter plot smoothing or LOESS (STL), into their trend, seasonal, and remainder components. The seasonal current density, which captures the photoresponse induced by the periodic light, was then used to train a LSTM network to predict the one-step-ahead seasonal current density. Results It is shown that the trained LSTM network is able to predict the seasonal current density and photoresponse in the BPVs to a high accuracy, using only lagged values of the current density and light status (on/off) as the predictive inputs. Mean absolute errors of 0.007, 0.014 and 0.013 μA m-2 were achieved on the training, validation and test data sets using a network of 35 neurones and a window size of 144. Errors were largest when light was switched on and, to a lessor extent, off. It is hypothesised that biofilm fluorescence yield may be an additional input used to improve predictions during light changes. This is an important first step to developing useful predictive models and optimisation algorithms for BPVs. McCormick, A. J. et al. Biophotovoltaics: Oxygenic photosynthetic organisms in the world of bioelectrochemical systems. Energy Environ. Sci. 8, 1092–1109 (2015). Tschörtner, J., Lai, B. & Krömer, J. O. Biophotovoltaics: green power generation from sunlight and water. Front. Microbiol. 10, 866 (2019). Longatte, G., Guille-Collignon, M. & Lemaître, F. Electrocatalytic Mechanism Involving Michaelis–Menten Kinetics at the Preparative Scale: Theory and Applicability to Photocurrents from a Photosynthetic Algae Suspension With Quinones. ChemPhysChem 18, 2643–2650 (2017). Okedi, T. I. & Fisher, A. C. Time series analysis and long short-term memory (LSTM) network prediction of BPV current density. Energy Environ. Sci. (2021). doi:10.1039/D0EE02970J Figure 1