A Capacity-Prediction Model for Exploration of Organic Anodes: Discovery of 5-Formylsalicylic Acid as a High-Performance Anode Active Material

Abstract

Development of high-performance organic energy storage is one of the important challenges in recent materials science. Molecular design and synthesis have potential for enhancement of the performances. Efficient exploration and design of the molecules are required in a wide search space. In the present work, a capacity prediction model for organic anodes was constructed on small experimental data by sparse modeling, a method of machine learning, combined with our chemical insights. The straightforward linear regression model facilitated discovery of a high-performance active material for organic anodes in a limited number of experiments. A recommended compound, 5-formylsalicylic acid (SA-CHO), showed one of the highest performances in recent works, i.e., a specific capacity of 873 mA h g −1 at 100 mA g −1 (sample number: n = 3) with rate performance and cycle stability. The model can be applied to explore organic anode active materials with higher specific capacity.

Capacity-prediction models for organic anode-active materials of lithium-ion batteries: advances in predictors using small data

“…aThe molecular structures of 1–54 are displayed in Scheme S1 in the ESI.bThe specific capacity refers to the training and test datasets in our previous work. 48,49 cThe specific capacity was measured in the present work (Fig. S1 in the ESI).dThe differences in the specific capacity are caused by the differences in the current density.…”

“…The former and latter values are the measured capacity in the datasets G1 and G2, respectively. 48,49…”

“…Model G2 was constructed using the training data (training dataset G2) containing 23 x n ( n = 1–5, 16, 18, 19, 21–29, 31–36 in Table 2) and 25 y (compounds 7 , 8 , 10 , 14 , 15 , 17 , 20–22 , 24 , 26–35 , 38–40 , 42 , 43 in Table 1, Scheme S1 in the ESI†) by SpM-S (Tables S3 and S4 in the ESI†). 49 As the specific capacity was measured at a current density of 100 mA g −1 to accelerate the collecting of y for the construction of models G2 and G3 (Table 1), the capacity ( y ) of some compounds was different from that used for model G1. In addition, compounds with specific capacity 0 ( 1–5 , 9 , 11 , 12 , 16 ) were removed to adjust the balance of the training data.…”

“…Although the positive correlation of x 2 ( E LUMO1 ) is not directly explained, the positive correlation of the LUMO levels was used as the results of ML in model G2. 49 In the present work, x 2 is also adopted as a descriptor in model G3. Further studies including a calculation study are needed to elucidate the correlation between the LUMO level and capacity.…”

“…10,14 Therefore, SpM-S provides straightforward, interpretable, and generalizable linear regression models using a limited number of descriptors. Our group constructed performance predictors for organic cathode-active and anode-active materials using SpM-S. [48][49][50] Although two predictors (G1 and G2) for the specific capacity of an anode were prepared in our previous work (Fig. 1a-d), 48,49 their prediction accuracy was not sufficient.…”