Artificial intelligence-assisted chiral nanophotonic design
(Nanowerk News) Chiral nanostructures can enhance the weakly inherent chiral effect of biomolecules and highlight an important role in chiral detection. However, the design of chiral nanostructures was challenged by extensive theoretical simulations and exploratory experiments. Recently, Zheyu Fang’s group proposed a chiral nanostructure design method based on gain learning, which can locate metal chiral nanostructures with sharp peaks in circular dichroism spectra and improve the chiral detection signal. This work envisions the powerful role of artificial intelligence in nanophotonic design.
Chirality is a fundamental physical property which means that an object or structure cannot be superimposed on its mirror image. It plays an important role in biomedical sensing, because enantiomers with opposite properties lead to very different biological effects. However, it is difficult to detect trace amounts of enantiomers, because chiral bio-molecules are quite small than optical wavelengths and present weak dichroism signals.
Chiral plasmonic nanostructures have been used to enhance the interaction between optical waves and chiral molecules. General dichroism technology is based on circularly polarized waves, namely the spin angular momenta (SAM) of photons. Chiral metasurfaces, metamaterials and nanoparticles can be used to enhance the sensing signal because their intrinsic circular dichroism (CD) signals can be sensitively influenced by chiral biomolecules. In recent years, dichroism technology based on orbital angular momentum (OAM) has been proposed and investigated, in which signal dichroism is defined based on the differential response spectra of opposing vortex modes. Plasmonic resonance structures with intrinsic chirality can also enhance dichroism signals based on OAM.
However, the interactions between chiral molecules and chiral nanostructures are very complex. Different biomolecules may require different nanostructures to achieve optimally enhanced dichroism signals. Therefore, the design of chiral nanostructures consumes enormous computational resources in iterative electromagnetic (EM) simulations. Artificial intelligence (AI) is emerging as a powerful tool in nanostructure design, which can handle more complex problems and larger-scale data compared to traditional optimization algorithms. AI has been used successfully in the design of metasurfaces, photonic crystals, integrated wavelength routers, etc.
In a recent paper published in Opto-Electronic Science (“Chiral detection of biomolecules based on reinforcement learning”), Zheyu Fang and colleagues proposed a chiral nanostructure design method based on reinforcement learning (Fig. 1), in which exploration of new nanostructures and model updating are carried out simultaneously.
The introduction of reinforcement learning improves the quality of the training dataset and reduces the number of EM simulations. Figure 1(a) presents the construction of the training dataset. Different nanostructures were randomly generated at the start and their optical responses were calculated by EM simulations. As shown in the figure. 1(b), the next step is to train several artificial neural networks (ANN) to obtain mapping relationships between the geometry of the nanostructures and their spectra. Then, a new structure was designed with a Bayesian optimization algorithm based on predictions from ANN (Fig. 1(c)).
ANN recognizes nanostructures with strong chirality, generates possible optimized structures, and reduces computational resources spent on weak chiral nanostructures. Figure 1(d) shows the process of updating the training dataset. For nanostructures whose optical response predicted by ANN is significantly different, the optical response will be calculated by EM simulation. ANN inaccuracies indicate that there are some similar structures in the current data set, so the simulated data is added to the training data set.
For nanostructures that are consistently predicted by different ANNs, there should already be similar structures in the training dataset. After data update, ANN is retrained. Therefore, ANN ensures that the nanostructures added to the data set have a strong chirality potential even though the initial data set was constructed randomly.
To carry out the proposed scheme, the chiral unit to be designed is parameterized by a 40×40 unit coded with 0 or 1, indicating that there is a cube of gold or air at that position. ANN achieves the mapping between the optical response spectrum and the matrix representing the geometry. Lastly, three chiral metasurfaces with different CD peak frequencies were designed. Chiral metasurfaces were also created and measured to validate the proposed method. The experimental results are in accordance with the design results. The frequency shift of the CD spectrum induced by chiral molecules was also measured using left-glucose and right-glucose solutions controlled by a microfluidic channel. The shift in the resonance wavelength between glucose enantiomers with opposite chirality reached 7 nm, indicating an increase in the sensitivity of the chiral molecule.
The proposed method improves the quality of the training dataset and reduces the number of electromagnetic simulations compared to the classical exploratory method. Chiral nanostructures with significant CD values and high chiral detection sensitivity have been successfully designed and fabricated. In addition to the chiral nanophotonic design shown, other optical properties can also be designed because the algorithm is universal for every physical meaning of the optical response. This work envisions promising AI applications in nanophotonic and electromagnetic design.