The Spring School gathered again a full house of researchers from nine different countries representing 17 different nationalities (!) and a wide range of different research interests. These included materials science, drug discovery, geology, machine learning, and mathematics while some participants’ main focus was in experimental methods. From experimentalists, it was particularly pleasing to learn that the reputation of computational chemists is no longer only “academic toil” but the results can really be useful and collaboration is sought after. The scope and approach was seen unique: “I’ve never been to an event that connects so many different approaches and relevant things”, commented one participant.
Networking with future colleagues
Discussions over the frequent breaks as well as during the poster session and sauna dinner indicated the added value of bringing together researchers with complementing skills and interests. It turned out that participants with prior experience on tools that were found to be useful to others exchanged contacts and prepared to work together after the School.
This year the School attracted also some machine learning experts. I believe it is interesting to see actual applications and use cases how people beyond the core of machine learning method developer community have adopted these tools. Also, based on the discussions overheard over coffee breaks, the machine learning specialists and those aspiring to use the methods did found each other.
Several European supercomputing centres, including CSC, promote international collaboration in high performance computing through the HPC-Europa3 programme. HPC-Europa3 funds research visits to nine European countries with access to computational resources. Several of the Spring School participants appeared interested in this opportunity.
Machine learning has secured a foothold in computational chemistry
Machine Learning has been a trending topic in many fields of science and beyond. This year Filippo Federici Canova and Yashasvi Ranawat from Aalto University updated their introduction to the method by focusing the tutorials to function as a seamless tool with the computational chemistry methods of the previous day. The tutorials were distributed as Jupyter notebooks, which offer a convenient interactive platform to explore interactively the methods and which the participants can set up also for themselves later. The tutorials interface with Open Source python libraries, like scikit-learn, but can be adapted also to make use of e.g. TensorFlow if needed.
The NOMAD Center of Excellence services were also introduced as an example of new infrastructure aimed at creating additional value of already existing computational materials science results as well as a way to share your own results and make them discoverable for other researchers – but also for yourself. How many of you have trouble finding your own simulation results completed even a year ago? The NOMAD CoE also offers state-of-the-art machine learning driven data analytics tools directly available through a web browser.
Advanced parallel hands-on tutorial on Excitation Energies and Excited State Densities given by Dr. Mikael Johansson. Picture: Atte Sillanpää / CSC.
Overall, the School agenda was tight and as we strive to introduce the most important methods of computational chemistry it was not possible to cover every topic at great detail. Although, some participants criticized this in the feedback, many also saw this compromise useful. This theme was found in all topics covering simulation methods.
Individual methods, like coarse grained molecular dynamics, were seen as islands in the sea of computational chemistry and the School was seen as an attempt to connect these islands thereby creating a more holistic picture of the different approaches to chemical problems. If you like, the School can be seen to provide a high-level map of the archipelago so you can choose which islands to look closer and spend more time to make sure if the method matches your needs.
The School materials, including presentation slides and hands-on instructions and input files are available on the School homepage. The material also includes links to further in-depth treatments for self-study.
Choose the right method for your topic
Ambitious research problems require selecting carefully the computational tools. As in life in general, there is no silver bullet, or just one simple answer, that works for everything. One of the original ideas of the School was to quickly present the theory and approximations behind molecular dynamics and electronic structure theory and then proceed to the applications showing the strengths and weaknesses of each individual method.
Dr. Luca Monticelli introduced the importance of choosing the right method for the problem at hand in his introduction to molecular dynamics methods on the first day. On the following days, we heard the same message again in the context of electronic structure theory and machine learning suggesting a more general statement: this method is not intrinsically better or worse than the other one, but this works better for these kinds of systems or properties, and vice versa.
It is dangerous to use the methods as black boxes. Instead, it is important that the researcher actively chooses which approximations to make i.e. where to sacrifice detail and where the details are required. In practice, for real world research problems some details must always be compromised over better sampling or larger model system size, so it is very useful to be aware of several different methods and their validity.
A snapshot of Dr. Mikael Johansson's Introduction to Electronic Structure theory slides, where he highlights the importance of choosing the right method and paving the way to the tour through myriad methods and layers of approximations.
Top picture: Prof. Ville Kaila explaining the philosophy on quantum chemical simulation approach on biochemical systems. Picture: Atte Sillanpää / CSC.