Research Fellows

Ala Taftaf

Early Stage Researcher at INRIA

Biography
Research
Outreach

I was born in Abu Dhabi at Arabic United Emirates.
I spent most of my childhood in Zarzis , a small coastal town in southeast Tunisia, just south of the Island of Djerba.
After high school, I moved to to the capital of Tunisia to study Computer Science Engineering at the National School of Engineers of Tunis (ENIT).
In parallel with my last year of Engineering I was enrolled in the first year of the master in �Information Processing and image analysis\".
I succeeded after that my internship, which was in Inria at Sophia Antipolis in France and I obtained by the end of the second year a double diploma, one from the University of Paris Descartes and the other from the National School of Engineers of Tunis.

My passion about mathematics and physics in addition to my computer science skills have pushed me to apply immediately to the 7th thesis offer of the About Flow Project.
My research addresses two of the most frustrating limitations of AD at the moment: the first one is the need to better take into account the high-level structure of the source, more specifically the presence of fixed-point solvers and the second is the need to adapt to popular programming paradigms, more specifically MPI-parallel architectures.

About Project Flow is a great opportunity for me not only to improve my communication skills, but also to discover other cultures, other ways of thinking and to share experiences With other ESRs. It will also help me to improve my spoken english during conferences and workshops while traveling across Europe.
I was always surrounded by a professional team knowing that my work is crucial for the project and this was always very inspiring for me. The work to be done will surely have a huge impact on my future career and life.

Objectives:

� develop advanced support for efficient fixed-point source-transformation
in Tapenade
� develop MPI parallel message support in Tapenade

Contribute to Work Packages

Modern optimization platforms in CFD rely heavily on accurate derivatives to find every possible improvement of a current design. Of paramount importance are the gradients, central to the adjoint models, but developing such an adjoint by hand is long and error-prone. On the other hand, Automatic Differentiation (AD) by source transformation can mechanically turn the source of any model into a new source that computes its derivatives. Specifically the "reverse mode" of AD is tailored to compute gradients efficiently.

However, the CFD models themselves become increasingly complex and sophisticated: Implementing their adjoint model and consequently their gradient remains a challenge. In the European Project AboutFlow, we are addressing two of the most frustrating limitations of AD at the moment: (1) the need to better take into account the high-level structure of the source, more specifically the presence of fixed-point solvers and (2) the need to adapt to popular programming paradigms, more specifically MPI-parallel architectures.

1. Taftaf, A. , Pascual,V., Hascoet, L. "Adjoints of Fixed-Point iterations". 11th World Congress on Computational Mechanics (WCCM XI). 20-25, July 2014, Barcelona,Spain.

I conducted my outreach activity at the Science Festival event at Nice, France in October 2015. I was a part of the INRIA stand in which three computers were provided.

My aim was to engage with people of different ages: kids, students, adults in order to enrich their scientific culture and make them understand the different aspects of my research, specially the automatic differentiation by source transformation and shape optimisation.

For the kids and old people who don't know the programming yet, my activity was to show them the basics of computer programming via a framework called "scratch" . I helped them to create their own programs by making simple algorithms that move some sort of "doll" provided by the tool. I also asked every kid who understood the activity to help by teaching other kids of his age. This way, not only a maximum number of children was engaged but also every participant has strengthened the basics he has just learned.

For the teenagers who have already a basic understanding of the programming, but they don't know the derivatives yet, the aim was to explain to them the basics of shape optimisation. To do so, I prepared for them three planes of different shapes and asked them to select the one they thought to be the fastest. I explained to them why one shape could be better than the other and how we could find the best shape using mathematics. I explained also that mathematical computations by hand are long and error-prone and that's why the researchers rely on the computer to do these calculations for them, for instance by implementing the different mathematical equations as programs.

For the teenagers and adults who know the basics of programming as well as the derivatives, the activity aims to explain to them the automatic differentiation by source transformation. As activity, I showed them a program that implement two mathematical equations and asked them to imagine a program that could compute the derivatives of these two equations. I explained to them that the generation of the program that compute the derivatives can be automatic thanks to the different automatic differentiation tools, for example "tapenade".

I showed them "tapenade" web site and explained to them how they could use it.

The outreach was totally successful. The kids were enjoying the programming activity and the parents promised to continue the experience with their children at home. The teenagers found the applied mathematics as well as the optimisation domains very interesting and they might probably continue their studies in these fields. Even the old people enjoyed the programming experience and realised that is time for them to enrich their knowledge about computer science.

Acknowledgements: Special thanks to Inria stand members : Thierry Vieville, Valerie Francois, Martine Olivi, Magali Martin Stéphanie Sorres, Arnaud LEGOUT, Alain Dervieux, Gautier Brethes, Pierre Alliez, Nicolas Douillet , Matthias Caenepeel, Stefano Casagranda and Konstantinos Mavreas.