Neural Networks for Mass Regression for HH $\rightarrow$ bb$\tau\tau$

The value of the Higgs boson self-coupling is predicted by the Standard Model, but it is still largely unconstrained experimentally and thus being actively studied at the ATLAS experiment at the LHC. The HH $\rightarrow$ bb$\tau\tau$ decay channel is one of the most sensitive probes for studying this property. Regressing the di-Higgs invariant mass ($m_{HH}$) for events in this channel and understanding its distribution are important to improving the determination of the Higgs self-coupling. This thesis discusses recurrent, dense, and mixture density neural network models for predicting $m_{HH}$ from other event variables. These new models demonstrate significant improvement in both accuracy of mHH predictions and ability to separate $m_{HH}$ distributions for different values of the self-coupling constant when compared to the currently adopted Missing Mass Calculator method for $m_{HH}$ estimation in this channel.