Developer:
Antonio Díaz Pozuelo (adpozuelo@uoc.edu)HDNNP is my UOC's (Universitat Oberta de Catalunya) final project degree (TFG).
HDNNP is a feed forward neural network (FFNN) that calculates the macroscopic properties (energy, pressure, conductivity, etc.) of an atom's system. It does so as follows:
-
It learns from the properties of certain sets of systems of atoms (positions in a time sequence and energy for a given temperature and density).
-
It predicts the macroscopic properties of another system of different atoms, under temperature and / or density, to those used for learning.
There are examples of atom's systems in "data" directory (50 boxes are included for each temperature data file).
HDNNP is designed to use the massive parallel paradigm for intensive computation like symmetry functions calculus or neural network epochs. Thus, HDNNP needs a NVIDIA's VGA with CUDA arquitecture which must support compute capability 2.0 or higher.
HDNNP is implemented with NVIDIA CUDA SDK 6.5 and it uses the IMSL library for parameters optimization. Therefore it implements a fortran wrapper for languages intercommunication.
-
Software:
- NVIDIA CUDA Compiler (nvcc).
- Intel Fortran Compiler (ifort) or GNU Fortran Compiler (gfortran).
- IMSL Fortran compiled library.
- GNUplot and cairo library.
-
Hardware:
- NVIDIA's VGA CUDA capable arquitecture which must support compute capability 2.0 or higher.
Download HDNNP application:
git clone https://github.com/adpozuelo/HDNNP.git
cd HDNNP
Compile (Makefile is ready for Intel Fortran Compiler):
make
Execute HDNNP application (there are execution examples in HDNNP.sh):
./HDNNP.sh