Organic semiconductors are a promising alternative to silicon-based semiconductors due to their lower cost. However, developing these semiconductors faces challenges in transferring electric charge across molecules. Traditional techniques for sampling and simulation are slow due to the vast chemical space. To address this, machine learning techniques have emerged as a solution for faster and more accurate results. However, a problem remains in transferring results from a source domain to a similar target domain.

In this thesis, we focus on predicting transfer integrals for organic semiconductors using a technique called "Domain Adaptation." We utilize a graph neural network architecture known for its effectiveness in molecular prediction tasks. We compare two models: a vanilla graph network that approximates transfer integrals and a domain adaptation graph network that aims to make predictions domain-invariant.

Through eight experiments, we demonstrate the effectiveness of this transfer learning technique when the distance between domains is not too large. However, we also identify limiting factors, such as reduced accuracy. We conclude by discussing future prospects, including the application of domain adaptation in Bayesian optimization problems within material science.