Ashwati Das

Rapid trajectory design in complex environments via machine learning strategies

Designing trajectories to balance design space trade-offs and shifting mission requirements in complex environments demands rapid iterations in the design process and swift responses during flight operations. For example, near term cis-lunar activities are constrained by varied technological capabilities as well as the need to coordinate diverse routes through space for differing cargo and crew transport systems. A flexible and robust trajectory design strategy, and generalizable approach to uncover potential trajectory concepts for broad mission types, including low-thrust and chemical systems is therefore integral to enabling efficient transport. In this investigation, globally attractive solutions are sought by exploiting machine learning techniques to automate exploration of the design space, identify potentially productive links, and exploit combinatorics to forge sequences for path planning. A successful framework is summarized in terms of four components: (i) Database generation - compilation of well-known dynamical structures to form a searchable volume. (ii) Accessible regions - establishing reachability within the searchable volume for a given thruster/engine capability. (iii) Automated pathfinding - exploiting machine learning techniques to determine the transport sequence solving for an efficient path. (iv) Convergence/optimization - once the transport sequence is determined as a globally efficient concept, it is optimized locally by more traditional numerical strategies.

Dominik Hauger

Modeling of normalized Reynolds stress tensor using deep learning algorithms with embedded Galilean invariance

To perform accurate simulations in fluid dynamics is a major research area in engineering. The math behind was deemed important enough that the Clay Mathematics Institute announced it as a millennium problem in 2000. Up to now, the equations describing viscous fluid flow, the Navier-Stokes (NS) equations, have not even been proven to always have a solution. Most engineering flows of interest are turbulent. Direct Numerical Simulations (DNS) solve the NS equations directly and resolve all scales of the turbulence but come with high computational cost and are therefore infeasible for most problems. To save computing time, state-of-the-art simulations solve the Reynolds Averaged Navier-Stokes (RANS) equations instead. Through the averaging in the RANS equations they must be complemented by a problem-dependent turbulence model. A new approach to develop a more general turbulence model is by using machine learning. In my research I perform supervised learning of a deep neural network in TensorFlow, a deep learning framework developed by Google. I use RANS data of film cooling studies as input and DNS data as goal to predict the Reynolds stresses directly. The long-term goal is to implement this turbulence model into an inhouse RANS code.

Veerappan Prithivirajan

On the determination of critical pore size in the fatigue response of additively manufactured IN718 via crystal plasticity

The inherent defects associated with materials produced by selective laser melting (SLM) limits their usage in safety-critical applications. In our work, a crystal plasticity (CP) based framework is developed to identify the critical porosity characteristics, which quantifies the limiting scenarios of fatigue crack initiation at a pore rather than the crystallographic features. 3D virtual microstructures are developed based on the characterization of SLM IN718 for use in the CP simulations. Damage indicator parameters, such as the plastic strain accumulation, elastic stress anisotropy, resolved shear stress and triaxiality, obtained from the CP simulations are used to identify the most probable locations of crack initiation. Pores are explicitly added to the microstructure instantiations in a systematic manner by varying the size, location, and proximity between pores. The critical pore size is defined as the size beyond which the location of crack nucleation transitions from crystallographic features to the pore vicinity, which is determined to be 20 μm in the material of interest with an average grain size of 48 μm. This work is beneficial in qualifying SLM materials given the natural porosity inherent to the manufacturing process, by reducing the number of fatigue experiments.

What is the Research Symposium Series?

The Research Symposium Series is a department-sponsored forum for graduate students and advanced-level undergraduates to present their research to a general audience.

The Research Symposium Series is designed to: 

• Facilitate the exchange of ideas and knowledge among faculty and graduate students.
• Provide opportunities for students to develop their technical presentation skills.
• Promote the research activities of the department to undergraduates and other interested individuals.

2018 Prizes

• $500, $300, $200 for best three presentations
• $150 for best undergraduate presentation
• $150 for best abstract

Questions about the Research Symposium Series may be directed to:
aaerss@ecn.purdue.edu
https://engineering.purdue.edu/AAE/Academics/StudentOrgs/aaerss
*Winners in the presentation category cannot compete in that category the following year. The same applies for winners in the abstract category.