Machine learning provides an important help for modelling experimental devices.

Machine Learning (ML) has become an important pillar in material science. It has helped the community in different areas such as material property analysis, nanomaterial analysis, discovering new materials, drug discovery and design, and quantum chemistry. It is being used in different manufacturing sectors that deal with various chemical-based industries ranging from electronics design and fabrication to daily gadgets.

A collaboration between a postdoctoral lecturer at the Thin Film Laboratory in the Physics Department, Faculty of Science, Suez Canal University and a research assistant at Wigner Research Centre for Physics has shown the importance of ML as a tool to reduce the cost of experimentation with different settings. They have used 1,4,5,8-naphthalenetetracarboxylic dianhydride (NTCDA) as it has many favorable merits for developing organic semiconductors.

A photodiode

The fabricated organic photodiode architecture as shown in Fig. 1 is in the form of Au/NTCDA/p-Si/Al. Their study introduces a hybrid heterojunction based on NTCDA/p-Si for photodetection applications.

Fig. 1: (a) Architecture of the fabricated photodiode, and (b) the molecular structure of NTCDA.

Different modeling techniques have been used in this study to reproduce the I-V curve of the fabricated device. As a result, the voltage and the incident intensity of the light – as shown in the opening figure – were used as the main features for the ML models and the predicted current is used to be benchmarked against the experimentally measured data, i.e., the output current from the fabricated photodiode.

Exploring more possibilities with models

This allowed the researchers to verify the correctness of their models and also explore different experimentation settings – since the models can generalize well enough – without the need of refabricating the device under different conditions or repeating the experimentation to check for the photodiode response at a certain voltage and a specific light intensity.

The experimental results show that according to the obtained values of spectral responsivity, linear-dynamic range, specific detectivity, signal to noise ratio, and response time for the fabricated photodiode under the influence of UV light of intensity 20 – 80 mW/cm², the current device is suggested for many optoelectronic applications especially for detecting UV light detection manner. These obtained results ensure that the optimum light intensity – where this detector works efficiently – is 50 mW/cm².

Benchmarking the models

The four models used in this study are K-Nearest Neighbor, AutoML approach based on TPOT framework, Feed Forward Neural Network, and a Continuous Variable Quantum Neural Network. The authors have explored different models to document the ML pipeline journey of the four models and indicate which is more suitable depending on the final target and the available resources.

In Table 1, you can see that the four models managed to perform very well on the testing dataset because the authors treated the modeling problem as a regression one rather than a fitting procedure. The table reflects the natural behavior of the KNN and the resulting ML pipeline from TPOT as the training error is 0.

The FNN or the Artificial Neural Network (ANN) managed to be the best model amongst other ML models. The QNN, despite having the lowest set of tunable parameters, only 40, and a single quantum mode (Qumode) managed to compete against the FNN with more than 1,000 trainable parameters and converge faster.

Table 1: The loss or mean squared error values for each model and the relative usability of each method according to the required knowledge for each one to be developed properly.

The more practical model

The usability column indicates the relative required experience for anyone to develop such models. In terms of practical software developments, the TPOT method is the easiest one to be used as it only requires having a basic knowledge of software programming and a solid understanding of different fitting metrics and criteria. The rest is completely automated by its internal API.

The KNN model requires a solid understanding of ML basics and how to prepare the data for such a distance-based algorithm. The last two algorithms require an expert level especially the QNN model. In terms of pure fitting criteria, we recommend the KNN model – with a higher number of k-folds for hyperparameter tuning – to be the go-to model for such situations.

The importance of machine learning

It is worth mentioning that the authors have used a Displaced Squeezed state to encode the classical features for the QNN and only 8 layers as shown in Fig. 2.

Fig. 2: The configuration of the quantum circuit used for modeling. The grouped gates represent a single layer of the QNN. The dashed box is repeated 8 times. The ‘r’ in the state preparation represents the magnitude of the squeezing gate and α is the displacement magnitude.

The study strongly suggests the importance of using machine learning techniques and especially using quantum computing in material science simulations. Moreover, it practically proved the universality of the continuous-variable QNN in a regression problem.

On the other hand, the fabricated organic photodiode proved to be very useful and resilient under different settings. The field of organic semiconductors still has a lot to give to us and the authors believe that the best way to boost it is by using ML to accelerate the development process.

Ahmed M. El-Mahalawy, Ph.D., Thin Film Laboratory, Physics Department, Faculty of Science, Suez Canal University, Ismailia, Egypt; and Kareem H. El-Safty, Wigner Research Centre for Physics, Budapest, Hungary

This article has first been published by the African Physics Newsletter - © American Physical Society, 2021