We’re happy to announce version 2.4 of PhiFlow, our differentiable simulation framework for machine learning. Among others, it now has improved support for sparse matrices, preconditioners and plotting: https://github.com/tum-pbs/PhiFlow/pull/116

The new features include:

  • Improved plots with additional recipes for bar charts and error bars.
  • Improved learning curve visualization with vis.load_scalars()
  • Decorator @math.broadcast to make functions compatible with Tensors.
  • Preconditioned linear solves (experimental) with ilu and cluster preconditioners
  • Improved support for sparse matrices
  • Additional math functions, such as soft_plus, factorial, log_gamma, safe_div, primal, is_inf, is_nan
  • Explicit device management with math.to_device()
  • Tensor unstacking to dict using **tensor.dim
  • Jit-compilable sparse neighbor search
  • Improved support for Φ-trees, added math.slice().
  • Broadcast string formatter using -f-f”…”