This award will enable an international collaboration for software development in computational chemistry between researchers in the United States and the United Kingdom. This international collaboration arose from a joint NSF/EPSRC workshop on Software Development for Grand Challenges in the Chemical Sciences held in Oxford, United Kingdom in June 2011. This award will provide travel for project team members to attend plenary and planning meetings, and for exchange visits to catalyze initial subprojects.
The project will support a collaboration exploring software challenges arising as mathematical models of potential energy surfaces become more accurate and also more complex. Outcomes of this award will include software for computational chemistry and improved international collaboration in support of sustainable software for science and engineering.
Molecular simulation lies at the heart of physically-inspired atomistic approaches to computational chemistry and biochemistry. The potential energy surfaces that describe basic molecular interactions are the algorithmic core of all molecular simulations, and together with the computer programs that implement them, they define the central cyberinfrastructure in this field. We sought to address the software challenges that arise when the mathematical models of potential energy surfaces become more accurate but algorithmically more complex, in order to create robust, sustainable software that will ultimately be disseminated to simulators world-wide. From the scientific point of view, empirical potential energy surfaces have been approaching a generational transition over the past several years, moving away from well-established but intrinsically limited fixed point charge models towards more intricate and expensive potential energy surface models that can allow accurate reproduction of a much broader range of reference data. The corresponding greater accuracy introduced by model improvements in short range repulsive forces, polarizability, charge transfer, or explicit quantum mechanical treatments open up several challenges in their manifestation as algorithms and software on current or emergent hardware platforms. These challenges were addressed by our UK and US consortium experts in molecular simulation, force field development, optimization and HPC and computer science, providing a unique combination of complementary skills. Travel funding was used two organize two face-to-face meetings in the US and the UK, and additional organization surrounding early pump priming work using cpu allocations on HECToR and XSEDE machines were applied for and successfully funded. Furthermore we organized weekly US-UK phone calls between US-UKparticipants. The major outcomes of the workshop produced a strong consensus that polarizable force?fields such as AMEOBA offer huge potential, however these haven’t been used extensively due to performance issues and complexity of software development, making simulation of real system untenable. This appeared to be the case for all the major MD software packages represented (AMBER, CHARMM, TINKER and OpenMM), while end users from the UK reiterated this point based on experience with the European code GROMACS and with codes based in research groups in the UK. Applying good software development skills with an emphasis on improved exploitation of HPC and accelerator technologies could therefore make a significant impact. Given that computational cost is a major barrier, the workshop believed an additional solution is to develop a hierarchy of PES models that alter the trade?off between accuracy and computational speed, and join into appropriate hybrid combinations to define a "sweet spot" for a given scientific application. We discussed the importance of software sustainability and in ensuring any software developments in any future proposal were available and applicable to the wider community. An exchange of ideas and practises in both the US and UK were discussed. Associated with this was a discussion on how to allow better exchange of software developments across multiple MD code suites and indeed between other quantum codes such as QChem, an important requirement if the model developments described above are to be fully exploited. The workshop participants were excited by the idea of developing a domain specific language (DSL) or Application Programming Interface (API) to provide a mechanism to allow users to specify familiar concepts without the need to handle the full complexity of the underlying hardware and/or software. This would provide a mechanism in which to abstract a problem and offer the possibility of better code exchange as well as allowing auto?tuning techniques to be exploited. A number of the talks identified critical problems that could benefit from the improved software and models discussed above, from both the UK and US. The workshop attendees were therefore confident that a significant range of potential beneficiaries are identifiable, again from both the UK and US. The need for these users to be actively engaged in the whole software development cycle was highlighted, using a ‘co?design’ process between software and applications, analogous to many current HPC projects (for example, the EU CRESTA project). The workshop also identified a strong need for US/UK collaboration in this area. The overlap in codes utilised by the UK and US is significant – the majority of the work in this area falls on five or six codes. Most of these are US based codes, with the exception of the European (Swedish) GROMACS code, which is utilised by both the UK and US delegates. The US strength in application codes is complemented by the UK’s acknowledged strength in computational science and parallelisation. Both communities bring software engineering and computer science expertise and have strong ‘end user’ communities providing a wide range of UK and US based beneficiaries.
