3 ms·

In science and engineering, it's common to hear of "Computational X" where X is "Plasma Physics", "Geodynamics", "Ocean Science", "Electrical Engineering", "Aer

In science and engineering, it's common to hear of "Computational X" where X is "Plasma Physics", "Geodynamics", "Ocean Science", "Electrical Engineering", "Aerodynamics", or any of a myriad of other disciplines. In these fields, the primary task is formulating Science/Engineering questions that can be effectively answered using computation, then writing the software to do so. There is also the more generic field "Computational Science and Engineering" (CS&E) [1], which is primarily the development of methods and software that is applied in many specific disciplines. I work primarily on the library development side of CS&E, working with many applications to understand their needs, often finding ways to satisfy those needs with better software reuse and better ultimate flexibility, then implementing any components that belong in libraries. Additionally, we do basic research by recognizing unexploited common themes and designing new algorithms that will benefit existing and potential users. There is currently a great deal of demand for Computational X, for CS&E through the postdoc level, and for CS&E in industry and national labs. It is recognized that a versatile practitioner of CS&E can effectively assist Computational X for many values of X, since there really is a lot in common. Once you've put in 10000 hours of CS&E, plus 1000 hours in a handful of disciplines, you should be able to get orientedy quickly in new applications. At the faculty level, support for CS&E is highly variable. At some universities, X departments have the view that all their faculty should "X first, computation second", while the mathematics department wants to see theorems and isn't overly concerned with following through to see the method applied to obtain a new science or engineering result. When hiring committees say "we need CS&E, but some other department should have the CS&E faculty", it drives a lot of talent away from teaching roles (e.g., into industry), perpetuating the relative lack of early education in CS&E fundamentals. On the other hand, the trend is for universities to have more direct interest in CS&E (sometimes recognizing it with a dedicated department or inistitute). It helps that computation stimulates many interdisciplinary collaborations, leading to more diverse funding possibilities. [1] We have a big meeting next week: http://www.siam.org/meetings/cse13/ http://www.siam.org/meetings/cse13/

pmahoney 14y ago
What sort of backgrounds do people working in this field have? More specifically, I have a bachelor's degree in materials science and currently work doing run-of-the-mill web development (were the typical background of my peers is a bachelor's in CS). What steps might I take to move into computational engineering?
- jedbrown 14y ago
  There are a couple journals dedicated to computational materials science. Two active fields are phase field models and molecular dynamics. Phase field models are mesoscale continuum (PDE) models based on the Cahn-Hilliard and Allen-Cahn equations, and are useful for things like predicting crystal grain growth and reorganization under strain, as well as fracture and many other uses. In phase-field modeling, the interesting physics is a consequence of the free energy functional, which is phenomenological and different for each material, so software should make it as easy as possible to turn specification of the free energy into an accurate and efficient massively parallel simulation. One such approach is outlined in this paper [1] from some colleagues. The resulting discrete systems are stiff and have positivity properties that must be preserved, so scalable implicit solvers (like multigrid) as well as methods for variational inequalities are necessary. In molecular dynamics, there are many parallel scalability and expressivity challenges. Fast algorithms for long-range interactions (such as the fast multipole method) lead to less regular computational, but huge asymptotic (O(n) versus O(n^2)). Since better algorithms become more important as our hardware allows us to keep scaling up the problem sizes, Fast methods need to be applied in more places. There is also a lot of activity in multiscale modeling and numerical homogenization, where explicit microscale models (often kinetic) are used to provide missing components of meso/macro-scale continuum models. [1] http://dx.doi.org/10.1016/j.commatsci.2011.07.028 http://dx.doi.org/10.1016/j.commatsci.2011.07.028

pmahoney 14y ago

What sort of backgrounds do people working in this field have? More specifically, I have a bachelor's degree in materials science and currently work doing run-of-the-mill web development (were the typical background of my peers is a bachelor's in CS). What steps might I take to move into computational engineering?

jedbrown 14y ago
There are a couple journals dedicated to computational materials science. Two active fields are phase field models and molecular dynamics. Phase field models are mesoscale continuum (PDE) models based on the Cahn-Hilliard and Allen-Cahn equations, and are useful for things like predicting crystal grain growth and reorganization under strain, as well as fracture and many other uses. In phase-field modeling, the interesting physics is a consequence of the free energy functional, which is phenomenological and different for each material, so software should make it as easy as possible to turn specification of the free energy into an accurate and efficient massively parallel simulation. One such approach is outlined in this paper [1] from some colleagues. The resulting discrete systems are stiff and have positivity properties that must be preserved, so scalable implicit solvers (like multigrid) as well as methods for variational inequalities are necessary. In molecular dynamics, there are many parallel scalability and expressivity challenges. Fast algorithms for long-range interactions (such as the fast multipole method) lead to less regular computational, but huge asymptotic (O(n) versus O(n^2)). Since better algorithms become more important as our hardware allows us to keep scaling up the problem sizes, Fast methods need to be applied in more places. There is also a lot of activity in multiscale modeling and numerical homogenization, where explicit microscale models (often kinetic) are used to provide missing components of meso/macro-scale continuum models. [1] http://dx.doi.org/10.1016/j.commatsci.2011.07.028 http://dx.doi.org/10.1016/j.commatsci.2011.07.028