Successful physics and performance studies of the ATLAS detector at the Large Hadron Collider rely on a large number of simulated events. The production of these simulated events with the precise detector description using GEANT4 is highly CPU intensive. With the large collision dataset expected to be collected by the ATLAS detector, the development of a simulation tool to reduce CPU requirements is imperative. During the LHC Run-1, a fast calorimeter simulation (FastCaloSim) was successfully used by ATLAS. FastCaloSim utilizes a parametrization of the energy response of particles at the calorimeter read-out cell level, taking into account the lateral shower profile and the correlation of the energy deposition among various calorimeter layers. The tool is interfaced to ATLAS digitization and reconstruction software and provides a calorimeter simulation approximately 500 times faster than GEANT4. An improved version of FastCaloSim is currently under development to further optimize the CPU and memory requirements and to improve the physics description. The new FastCaloSim implements machine learning techniques, such as principal component analysis and neural networks to optimize the amount of information stored in the ATLAS simulation infrastructure. These techniques improve the physics modeling and enhance the performance by reducing the I/O time and the memory usage during simulation. In this talk, the new FastCaloSim parameterization will be described, its performance will be quantified and its physics applications discussed.