Reaserch Works

Implementation of the confusion scheme on snapshots of ultra-cold quantum gas experiment to detect the topological phase transition of the Haldane model

To implement the confusion scheme on this dataset a few modification has been made to the original method, first as the data is high dimensional I've used a deeper network with convolutional layers as opposed to the shallow fully connected network suggested by the paper. Secondly the Haldane model in the phase φ=-π/2 has two critical frequencies, theoretically the Haldane model is topologically trivial in ω <6.1 and ω> 6.8 and topological, C=-1, in between. So we have a double phase transition on our hand and the universal W shape here has two peaks in the middle showing two critical points.

Discriminative Cooperative networks on snapshots of ultra-cold atoms

This is a try to apply the DCN method to detect topological phase transition of the Haldane model from snapshots of the ultra-cold atoms experiment. The dataset is the same as the one used in the paper Unsupervised machine learning of topological phase transitions from experimental data and you can access it from here .

I suspect the problem is either

1. Data being high dimensional and DCN method has been implemented only on low dimensional (80d) array of entanglement spectrum and one key difference from confusion method is the interpolation of data in continuos parameter space and in their implementation they have used a linear interpolation that simply doesn't work on our images. I tried to overcome this issue by implementing a generative model with variational autoencoder with question neuron as suggested in this paper and has been utilized to remove micromotions from this dataset as well. I added two more question neurons for frequency and phase which the snapshots has been taken in, and in this way you can control the generated image parameters but there is no guarantee that the generated snapshots contain meaningful physical properties (at least I couldn't find a way).
2. One issue that hasn't been addressed in the DCN method is that if the guesser model is trying to minimize the loss function of learner model from the confusion method the global minima of the loss function is where all labels are the same (the boundary peaks of the universal W shape) and not the critical point. For overcoming this issue I tried to make the guesser model to propose the right critical point by rewarding the higher variant labels but it just made the model to not learn anything.

Generating Snapshots of Ultracold Quantum Gas Experiment with Variational Autoencoders

This is a try to generate new samples from this dataset by adding two more question neurons to the original model that has been used in order to remove the micromotions phase. The idea is that if one can change the micromotion of sampled images with experiment why can't you just generate new images in the desired parameters (with specific frequencies and phases in our case). The problem is that I couldn't find a clear way to evaluate the generated images, and it raises the question wether the generated snapshots preserve meaningful physical properties. If you have any suggestion regarding this I would be happy to discuss this.

Reformulating the detection of quantum phase transition as a deep reinforcement learning problem

I defined the environment as a set of nodes living in the parameter space, presenting the phase diagram of the many-body physical system under study. The agent can take action to change the coordinates of these nodes to obtain the physical system's phase diagram. The rewarding system for this environment is defined as the performance of a shallow neural network trained with the suggested phase diagram by the agent as an indicator of how well it is doing (confusion scheme). Here I used the QLearning method which requires a discrete action space and by using SAC we can use a continous action space.

Soft Actor-Critic Learning Physics

Coming soon...