Significant theoretical progress has been made in recent years on the problem of three neutral atoms colliding at ultracold temperatures. Much of that progress has relied on computational physics, and there is now a unique opportunity to substantially advance the computational solution of the ultracold few-body problem. This opportunity comes from an unlikely source: computer video gamers. The computational power that has been built into computer video cards to drive games is enormous, promising orders of magnitude improvement over computing on typical workstations. Moreover, recent developments have made it much easier for the high performance computing community to tap into this power, and that is what is proposed here. At the same time, several improvements are proposed to both the underlying representation of the problem and the numerical algorithms employed. Taken together, this three-pronged attack on the problem can be expected to render currently intractable problems in ultracold few-body physics tractable. In turn, these advances in computational physics can be expected to further our theoretical understanding of these problems.
The goal of this proposal was to improve our computational solution of the ultracold few-body problem by attacking three fronts: the mathematical representation of the problem, the computational algorithms used, and the computer hardware. We identified the most promising possibilities in each category to see if they improved the performance of our code in a tangible way. Testing these possibilities required understanding the new method, deriving the details of implementing it for our problem, writing and debugging the necessary computer code, and finally performance testing the resulting code against our existing methods. While one of the new mathematical representations we tried did not offer any improvement, another did — although we could not fully test it before the grant period ended. In contrast, we were able to reduce the computer memory required by a factor of two or more (in some cases). Since our calculations require up to tens of gigabytes, this is a substantial improvement. The same modification also reduced the computational time. Finally, the highest risk aspect of this Project — trying to adapt graphical processing units (GPUs) to benefit our calculation — did not pay off. If it had, we might have realized a tenfold or more reduction in the computation time. Unfortunately, GPUs are very specialized computer hardware that are not appropriate for every problem. Such high-risk, high-reward hardware solutions must be periodically re-investigated, and they may yet benefit our research. Nevertheless, our experience with GPUs has already allowed us to help several colleagues get started using them for problems that will benefit from them. A former group member has even taken GPU experience from our group to her new job in medical imaging. Before the postdoctoral associates carrying out this work could actually try to implement any changes to our codes, they had to learn to use our codes and understand them. The reason for this is that our tasks are sufficiently specialized that we must write the computer code from scratch rather than using existing software. The postdoctoral associates thus undertook "training projects" using the code. These training projects produced important results for few-body physics and resulted in high-visibility publications. One settled a decades-old controversy, and the other discovered a new class of three-body molecules. These results will thus impact atomic physics, molecular physics, nuclear physics, and chemistry.
