Differentiable Thermodynamic Phase-Equilibria for Machine Learning