Machine learning with knowledge constraints for process optimization of open-air perovskite solar cell manufacturing

•ML accelerates open-air process optimization for perovskite solar cells•With a budget of 100 conditions, 18.5% device efficiency is achieved•Researchers’ domain knowledge can be incorporated into ML optimization•Benchmarking results show an advantage over traditional methods Perovskite photovoltaics (PVs) have achieved rapid improvement in the past decade power conversion small-area lab-scale devices. However, successful commercialization still requires development low-cost, scalable, and high-throughput manufacturing techniques. Machine learning (ML) materials science engineering has been developed recent years, it readily used to accelerate scale-up. We demonstrate Bayesian framework that allows incorporation researchers’ ML-guided loop. In case optimizing cells by spray plasma processing (RSPP) technique, proposed enables faster comparison with other conventional researcher-driven design-of-experiment methods. Although shown RSPP, broadly accelerated technologies PVs. Developing scalable technique high-dimensional parameter space. Herein, we present machine (ML)-guided sequential cells. apply our methodology fabrication. With limited experimental screening demonstrated as best result from fabricated RSPP. Our model enabled three innovations: flexible transfer between processes incorporating data prior probabilistic constraint, both subjective human observations insights when selecting next experiments, adaptive strategy locating region interest using before conducting local exploration high-efficiency Furthermore, virtual benchmarking, achieves improvements budgets than design-of-experiments Metal halide perovskites are efficient absorbers compatible low-cost solution promise emerging thin-film photovoltaic (PV) technology. Scaling up fabrication currently one critical research areas technology on path commercialization.1Li Z.C. Klein T.R. Kim D.H. Yang M. Berry J.J. van Hest M.F.A.M. Zhu K. Scalable cells.Nat. Rev. Mater. 2018; 3: 1-20Google Scholar, 2Perini C.A.R. Doherty T.A.S. Stranks S.D. Correa-Baena J.-P. Hoye R.L.Z. Pressing challenges photovoltaics—from atomic module level.Joule. 2021; 5: 1024-1030Google 3Li D. Zhang Lim K.-S. Hu Y. Rong Mei A. Park N.-G. Han H. A review scaling cells.Adv. Funct. 31: 2008621Google Scholar Despite success >25% academic labs spin coating,4Almora O. Baran Bazan G.C. Berger C. Cabrera C.I. Catchpole K.R. Erten-Ela S. Guo F. Hauch J. Ho-Baillie A.W.Y. et al.Device performance materials.Adv. Energy 11: 2002774Google Scholar,5Yoo Seo G. Chua M.R. T.G. Lu Rotermund Y.-K. Moon C.S. Jeon N.J. al.Efficient via improved carrier management.Nature. 590: 587-593Google this method not line. Recently, Rolston al.6Rolston N. Scheideler W.J. Flick A.C. Chen J.P. Elmaraghi Sleugh Zhao Woodhouse Dauskardt R.H. Rapid modules.Joule. 2020; 4: 2675-2692Google because potential achieving PV modules at cost ∼$0.2 /W. addition cost, main spray-deposition-based its ultrahigh throughput mechanical properties thin film methods, such blade coating, slot-die roll-to-roll printing. For new (including RSPP), typically takes months years achieve control reproducibility scale, several estimate upper One key there many parameters co-optimize1Li Scholar,6Rolston Scholar—e.g., precursor composition, speed, temperature, head/nozzle height, curing High-throughput experimentation introduced explore regions space.7Ahmadi Ziatdinov Zhou Lass E.A. Kalinin S.V. metal perovskites.Joule. 2797-2822Google 8Zhao Xu Z. Sun Langner Hartono N.T.P. Heumueller T. Hou Elia Li al.Discovery temperature-induced stability reversal robotic learning.Nat. Commun. 12: 2191Google 9Du X. Lüer L. Wagner Osterrieder Wortmann Vongsaysy U. Bertrand al.Elucidating full OPV utilizing robot-based platform learning.Joule. 495-506Google 10Saliba Polyelemental, multicomponent semiconductor libraries through combinatorial screening.Adv. 2019; 9: 1803754Google 11Zhao Luo Kasian Kupfer Liu B. Zhong al.A bilayer polymer structure planar 1,400 hours operational elevated temperatures.Nat. Energy. 2022; 7: 144-152Google some cases, problem nearly impossible solve brute force or even sophisticated design experiments (DoEs). Sequential e.g., (BO), emerged effective strategies wide range chemical reaction synthesis12Burger Maffettone P.M. Gusev V.V. Aitchison C.M. Bai Wang Alston B.M. Clowes R. mobile chemist.Nature. 583: 237-241Google 13Granda J.M. Donina Dragone V. Long D.L. Cronin Controlling organic synthesis robot search reactivity.Nature. 559: 377-381Google 14Shields B.J. Stevens Parasram Damani Alvarado J.I.M. Janey Adams R.P. Doyle A.G. tool synthesis.Nature. 89-96Google material optimization.15Gongora A.E. Perry W. Okoye Riley P. Reyes K.G. Morgan E.F. Brown K.A. autonomous researcher design.Sci. Adv. 6eaaz1708Google 16MacLeod B.P. Parlane F.G.L. Morrissey T.D. Häse Roch L.M. Dettelbach K.E. Moreira Yunker L.P.E. Rooney M.B. Deeth J.R. al.Self-driving laboratory discovery materials.Sci. 6eaaz8867Google 17Mekki-Berrada Ren Huang Wong W.K. Zheng Xie Tian I.P.S. Jayavelu Mahfoud Bash al.Two-step optimized nanoparticle synthesis.npj Comput. 55Google 18Attia Grover Jin Severson Markov T.M. Liao Y.H. M.H. Cheong Perkins al.Closed-loop fast-charging protocols batteries learning.Nature. 578: 397-402Google 19Lookman Balachandran P.V. Xue Yuan Active emphasis sampling uncertainties targeted design.npj 21Google 20Nikolaev Hooper Webber Rao Decker Krein Poleski Barto Maruyama Autonomy research: study carbon nanotube growth.npj 2016; 2: 1-6Google 21Balachandran Kowalski Sehirlioglu Lookman Experimental high-temperature ferroelectric guided two-step 1668Google 22Ling Hutchinson Antono E. Paradiso Meredig High-dimensional data-driven well-calibrated uncertainty estimates.Integr. Manuf. Innov. 2017; 6: 207-217Google 23Rohr Stein H.S. Guevarra Haber J.A. Aykol Suram S.K. Gregoire Benchmarking acceleration learning.Chem. Sci. 2696-2706Google 24Erps Foshey Luković M.K. Shou Goetzke H.H. Dietsch Stoll von Vacano Matusik Accelerated 3D printing multiobjective optimization.Sci. 7eabf7435Google 25Tran Tranchida Wildey Thompson A.P. Multi-fidelity machine-learning quantification design: application ternary random alloys.J. Chem. Phys. 153074705Google BO work well problems under 20 variables26James Miranda PySwarms: toolkit particle swarm Python.J. Open Source Softw. 433Google 30 variables algorithm modifications.27Wang Gehring Kohli Jegelka Batched large-scale spaces.in: Proceedings 21st International Conference Artificial Intelligence Statistics. AISTATS, 2018: 745-754Google Scholar,28Harris S.J. Harris D.J. Failure statistics commercial lithium ion batteries: 24 pouch cells.J. Power Sources. 342: 589-597Google Therefore, chosen sequential-learning-based current RSPP Current reports studies classical two common drawbacks: (1) no direct channel incorporate information previous relevant (2) inflexibility adapt qualitative feedbacks iterative On hand, sometimes significant amount learn what already become apparent researchers.29Ziatdinov M.A. Ghosh Physics makes difference: active augmented Gaussian process.Mach. Learn. Technol. 3015022Google these drawbacks could discourage researchers adopt tools their utilized iterations. Previous useful sources planning. An acquisition function “decision-maker” produce plan round. constraints (introduced Gelbart al.30Gelbart Snoek unknown constraints.Preprint arXiv. 2014; (1403.5607)Google Scholar). aim optimization, al. density functional theory (DFT) calculations phase constraint composition improve stability, avoiding compositions susceptible segregation.31Sun Tiihonen Oviedo Thapa Goyal Batali fusion approach optimize compositional perovskites.Matter. 1305-1322Google Simple visual assessment thickness, color, structural defects powerful provides additional guidance intelligent optimization. words, if experienced identifies low-quality film, subsequent longer necessary. Defining way framework. work, develop (PCE) target variable. As illustrated Figure 1, iteratively learns process-efficiency relation suggests optimal target. general framework, start planning conditions model-free initial Then, method, PCE measured simulator standard testing (STCs). PCEs, train regression relation, subsequently predict (and prediction uncertainty) unsampled regions. Finally, evaluated together information, therefore, round planned. By including (i.e., quality related study), maps out space tends avoid less promising based observational film-quality data. Hence, suggestions focus most within To capability approach, consider six input absorber layer process. aimed exceed produced First, obtained reaching five Second, describe how multiple were fused constraint. Third, analyze learned relationships efficiency, extracting generalizable insights. Fourth, benchmark factor enhancement against DoEs simulations, demonstrating excellent fewer conditions. 2A plots inspection PCEs batches consisting time enable iteration feedback suggest parameters. The highest each condition (the dark-green dots 2A) was algorithm. addition, dataset 45 (Figure 2C), defined top performer above 17% good exceeding 15%. Thus, had 1 6 performers. Among 85 BO-guided (excluding films did pass inspection), >15% (good performers), 10 >17% (top performers). rate therefore 47% performers 12% contrast, among Latin hypercube (LHS)-guided 50 2B), found only (12% rate) (2% rate). champion best-in-our-lab LHS never reached 18% PCE. According paper spray-deposited cells,32Bishop J.E. Smith Lidzey D.G. Development spray-coated cells.ACS Appl. Interfaces. 48237-48245Google comparable highest-efficiency devices deposition open air (18.5%)33Su Cai Ye Ni ink ultrawide window ambient air.ACS 3531-3538Google N2 glove box (19%).34Ding Q. Ge Q.Q. Ma J.Y. B.Y. Y.X. Mitzi D.B. J.S. Fully air-bladed photovoltaics.Joule. 402-416Google Visual done after depositing rated evaluating uniformity, pinholes. ratings intended very conservative so lowest-quality “tossed.” This evaluation step typical processing. Example high-quality S1. confirm validity criteria, those films. confirmed low values all below 13.5% average 7.8%), corresponding distribution S2. validation confirms skipped. When training learning, function. Note first batch films, which further demonstrates sampled more effectively time. 3 shows study. Because six-dimensional difficult visualize, speed substrate temperature illustrations. similar plot pair analysis. generated objective functions. contains primary plotting relationship 3A). combination layers included (constraint 1; 3B) 2; 3C). resulting value determined selection criteria quality). These plotted Figures 3D–3F provide namely goal higher performance. raw confidence bound (UCB) Probabilistic calculated probability good-quality 2 above-average dataset. Both functions scaled reduce impact prevent modification being too harsh. More specifically, binary outcome either 0 (fail) (pass). built interpolate (see 3E) sequentially converted passing assessment. 1) then softened 0.5–1 multiplied 3B). See supplemental section 1.2 mathematical definition means weigh times important any bias quality. (that 1B) acquired during preliminary setup. also use instead beyond considered herein slightly modified. example, nozzles different work. believe essential transferred affected strongly modifications), equivalent required subtle representing directly adding training. 4 visualizes evolved began converge Including LHS, rounds conducted 1–3 followed affect 1. higher-temperature higher-speed observation consistent 3E 3F. Due (or total 5 rounds), opted final Aiming PCE, smaller around predicted model. method,26James Scholar,35Kennedy Eberhart Particle optimization.Proceedings ICNN'95—International Neural Networks. 4. 1995: 1942-1948Google global optimum given Table reason change (GP) tendency “smooth out” features response surface 5. finding science,17Mekki-Berrada where GP fit smooth overfitting small number points. increase value, ringfencing identify conditions), rather specific point probability. Within probability, combined few techniques balance exploitation exploration. consisted model-predicted condition, nearest neighbors exploitation, visualized trained relations projecting 2D pair-wise contour plots. plot, × grid. every variables), remaining four 200 times, interested maximizing took maximum manifolds reduced variables, 15 possible examples dimension-reduced manifolds. indicates = 140°C duty cycle 20% >16.5% fully optimized. regressor lower experimentally ground truth) due error difference BO, but balancing compensates error. errors S4. manifold inform correlations efficiency. projected suggested onto plots, helps interpret decision-making mistakes Some led negative correlation gas flow/duty (Figures 5A 5B). Since certain dose convert crystalline perovskite, indicated increased offset energy delivered Additionally, correlative trends (consistent experimentation) observed. flow correlates 5C), since constant should form desired thickness. simulated “virtual” “teacher” benchmarking DoE gradient boosting decision trees, 2B). 6A. needs ensure predictions follow monotonic truth actual experimentation). following al.,36Häse Aldeghi Hickman R.J. Christensen Liles Hein Aspuru-Guzik Olympus: noisy experiment planning.Mach. 2035021Google spearman coefficient metric assess strength ground-truth efficiencies. approximates 0.93, mod