NPE for $\beta$-pic b

Using Neural Posterior Estimator to retrieve the orbital parameters of the $\beta$-pictoris b planet given the direct imagery observations.

---

Dependencies

• orbitize
• lampe
• zuko
• h5py

---

Generate Datasets

The script generate.py creates training, validation and test datasets in HDF5 format. These datasets are used to train the neural posterior estimator.

Here's how to run the script:

python generate.py --size [size] --name [name]

---

MCMC

The script mcmc.py performs Markov Chain Monte Carlo (MCMC) to estimate the orbital parameters. Takes more than 24 hours

python mcmc.py

---

Train the model

python generate.py --size [size]

Should be the same size as was generated

---

Results :

The corner plot obtained with the two methods. MCMC took around 27 hours to run and the Neural posterior estimation model about 1h on a GTX 1080 Ti for a trainingset of size $2^{18}$.

We can also visualise the confidence interval produced by the two models with the real observations

MCMC

NPE