Inverse design of chiral functional films by a robotic AI-guided system

Artificial chiral materials and nanostructures with strong and tuneable chiroptical activities, including sign, magnitude, and wavelength distribution, are useful owing to their potential applications in chiral sensing, enantioselective catalysis, and chiroptical devices. Thus, the inverse design and customized manufacturing of these materials is highly desirable. Here, we use an artificial intelligence (AI) guided robotic chemist to accurately predict chiroptical activities from the experimental absorption spectra and structure/process parameters, and generate chiral films with targeted chiroptical activities across the full visible spectrum. The robotic AI-chemist carries out the entire process, including chiral film construction, characterization, and testing. A machine learned reverse design model using spectrum embedded descriptors is developed to predict optimal structure/process parameters for any targeted chiroptical property. A series of chiral films with a dissymmetry factor as high as 1.9 (gabs ~ 1.9) are identified out of more than 100 million possible structures, and their feasible application in circular polarization-selective color filters for multiplex laser display and switchable circularly polarized (CP) luminescence is demonstrated. Our findings not only provide chiral films with the highest reported chiroptical activity, but also have great fundamental value for the inverse design of chiroptical materials.

Supplementary Table 1.Estimated numbers of parameter combinations before and after the clustering-screening.  16) and the number of epochs (50).To connect the generator with the forward network, the first number of the generator's output, who represents the species of dyes, is converted into a one-hot vector, and then multiplies with a 10×121 matrix where each row represents the absorption spectrum of one dye.As a result, the species of dyes are transformed into the corresponding absorption.
Supplementary Table 2.The recipes of the Fig. 2c

Supplementary Fig. 5 |
Results of the clustering-screening for transparent films.
(a) Histogram of LD and T0 for transparent films.(b) after hierarchical clustering, transparent films form four clusters with one high performance (highlighted in red shading) and one low performance (highlighted in blue shading).(c) Phigh and Plow of parameter values for transparent film in the first round screening (n=39).(d) Phigh and Plow of parameter values for transparent film in the second round screening (n=30).

Supplementary Fig. 12 |
Thickness dependence of model predicted chiroptical activity of the composite films.The model predicted dissymmetry factor gabs values of 10 dyes at different thickness under the condition of strain 100%, greyscale 5 and angle 45°.The distance between peaks and troughs of gabs spectra decreases with the increment of thickness.Supplementary Fig. 13 | Greyscale dependence of model predicted chiroptical activity of the composite films.The model predicted dissymmetry factor gabs values of 10 dyes at different greyscale under the condition of thickness 80 μm, strain 100% and angle 45°.With the increment of greyscale, gabs increase first and then decrease, while the waveform of gabs spectra keeps largely unchanged.Supplementary Fig. 14 | Overall architecture of the reverse model.A generator produces a random set of parameters.After being restructured into spectrum embedded descriptors, they are used as input parameters for the forward prediction model, which produces a circular dichroism (CD) spectrum based on the machine learned quantitative structure-spectrum-activity relationship (QSSAR).The key spectral features are compared with user-specified target properties.The loss (difference between predicted properties and target ones) is minimized by recursive optimization based on the Adam algorithm.Supplementary Fig. 15 | Schematic illustration of the generator in the reverse model.The generator is a fully connected neural network consisting of a 20-nodes input layer, a 5-nodes output layer, and 3 hidden layers whose numbers of nodes are 40, 80, and 20, respectively.All settings of the generator are the same as the forward network except the batch size (

Table 3 .
about experimental realization of The comparison of the dissymmetric factor gabs. (The theoretical limit of dissymmetric factor is ±2.)

Table 4 .
The comparison of the color gamut for the polarization dependent color switching.

Table 5 .
The comparison of the luminescence dissymmetric factor glum for the perovskite based circularly polarized luminescence.(The absolute of the idea dissymmetry factor is 2).Supplementary Fig. 31 | The circularly polarized photoluminescence (CPPL) spectra of eight perovskite quantum dots.Experimental realization of inverse design of filters to maximize left-handed circularly polarized photoluminescence (LCPPL), instead of Right-handed circularly polarized photoluminescence (RCPPL) shown in the main text.