Page 32 - Read Online

Wang et al. Energy Mater. 2026, 6, 600064 Page 25 of 34

Table 3. Representative machine learning-driven optimization strategies for perovskite solar cells
Experimental
Target Optimized
Dataset sources Algorithms Descriptors validation Refs.
properties additives/parameters
results
Training set: 63 data points Molecular weight, C
comprising 26 different atom, H atom, N atom, S
molecular additives atom, O atom, hydrogen
extracted from published bond donor, hydrogen
literature; virtual screening RF bond donor acceptor, PCE 5-CP PCE = 21.4% [41]
pool: 141 candidate rotatable bond,
molecules sourced from the complexity, TPSA, Ipc,
PubChem database MolLogP
14 experimental
processing factors 2-phenylethylammonium
generating 22 features iodide (PEAI) as HTL
2079 data points extracted including precise passivator, cuprous
from 1,149 experimental CatBoost stoichiometric ratios PCE thiocyanate (CuSCN) PCE = 25.01% [162]
articles published between precursor solutions doped Spiro-OMeTAD and
2013 and 2023
transport layer materials 4-chlorobenzenesulfonate
and various passivating as precursor additives
additives
Ground state property
Light gradient descriptor set (MDS-GS,
boosting containing 23
machine
127 data points comprising descriptors), the
experimentally measured (LGBM), absorption spectrum
Zn-porphyrin-sensitized artificial neural property descriptor set PCE D58-DP-ZnP-A44 PCE = 10.5% [164]
solar cells from peer network (ANN) (MDS-ABS, containing 19
reviewed literature sources and descriptors), and the
convolutional electron transfer property
neural network descriptor set (MDS-ET,
(CNN)
containing 5 descriptors)
An initial extremely limited
dataset comprising only 129 The Co-PAS Continuous high
known perovskite additives framework dimensional latent
extracted from literature integrating a vectors generated by the
which was subsequently MSC for robust JTVAE combined with key
utilized to systematically structural data molecular properties PCE BTN PCE = 25.20% [165]
screen a massive library of partitioning and including donor number
250,000 unknown a pre-trained dipole moment and
molecules randomly drawn JTVAE hydrogen bond acceptor
from PubChem count
Nanopatterning depth of PCE = 17.338%,
120 datasets collected from mesoporous-titanium
Jsc, Voc, fill
previous experiments (108 dioxide (mp-TiO 2 ), weight factor (FF), 127 nm nanopatterning Jsc = 22.877
-2
for training, 12 for RF depth and 0.10 wt% of mA cm , Voc = [166]
validation) percentage (wt%) of PCE PCBM 0.963 V, FF =
PCBM 78.7%
Experimental data generated Gaussian
in-house, starting with initial process
J-V curves of devices doped regression KI doping concentration,
with 0%, 4%, 8%, and 10% (GPR) equipped voltage PCE 3% KI doping PCE = 20.91% [167]
potassium iodide (KI), and with a squared
iteratively updated with new exponential
experimental rounds kernel
Multiobjective
bayesian
In-house generated data via optimization Volume ratios of three
a high-throughput (BO) utilizing precursor solutions
experimental platform, the Gaussian including pristine MAPbI 3 Peak PCE, 0.14 M MACl and 0.06 M PCE~11%, ΔPCE
initial dataset comprised 81 Process and solutions containing ΔPCE AVAI [168]
fully fabricated solar cells surrogate 5-ammonium valeric acid ~0.7%
exploring 21 ternary models and the iodide (AVAI) or MACl
compositions qNEHVI additives
acquisition
function