Seminar | ISE Graduate Student Colloquiums: Navneet Kumar & Sandeep Srinivasan

144 Baker Systems
1971 Neil Avenue
Columbus, OH 43210
United States

Title: Predicting stress hotspots in a microstructure using convolution neural network

Presenter: Navneet Kumar

Committee: Dr. Michael Groeber (advisor), Dr. Steve Niezgoda, Dr. Farhang Pourboghrat, Dr. Subhadeep Paul


Fracture in a material is mainly driven by the localization of response due to the applied stress and strain. These regions of high stresses are referred as hot spots which ultimately may lead to crack initiation. Factors leading to these hot spots depends on morphology of local features, their relative crystallographic misorientation and elastic anisotropy, and their preferred orientation with respect to the local state. Since a lot of factors may be responsible for the development of hot spots, it is advantageous to develop a method to identify these high stress regions using microstructure images. These images can subsequently be used to run an image-based machine learning algorithms such as convolution neural network (CNN).  The synthetic microstructures are generated in DREAM.3D that has its elastic response modeled using a spectral technique based on fast Fourier transforms. These microstructures are then cut into small patches of images which forms the training, validation and testing sets. The patches are cut in such a way that the center region contains the hot spot. The patches are then labeled based on the hot spot present in the center region, 1 for the hotspot contained patch and 0 for a non- hot spot patch. The CNN is trained on these images which is then later able to predict which region of a microstructure is susceptible to forming hot spot on an unseen dataset. 

The project is divided in two parts. The first part deals with the case where the material undergoes single thermal cycle of the particulate structure. To illustrate that the CNNs are capable of performing better than predicting at random chances, we vary the size of image patches, the architecture of the CNNs and the material properties in the simulation. The image sizes used are 16X16, 32X32, 64X64, 96X96, 128X128 and 256X256. The architecture of the CNN consists of blocks with each block consisting of a convolution layer, ReLU activation layer, a batch normalization layer and a max pooling layer. We vary the number of convolution layers from 1 to 4 in each block. The total number of blocks used are up to 4. The total convolution layers range from 6 to 96 covering a wide range of possible combinations. We try to come up with the optimal image size and adequate number of blocks with given convolution layers which gives best result.

The second part of the project discusses the case where the material undergoes multiple thermal cycles leading to accumulation of stresses at multiple locations in a material. The crystallographic and geometric descriptor such as Schmidt factor, Euler angles, grain shape parameter etc. play an important role in determining a region as hotspots. The significance of features is determined by machine learning algorithms such as linear regression and random forest which are then fed into CNNs to predict hotspots. We again vary the architecture of the CNNs and the material properties to demonstrate that the CNNs perform better than predicting at random chances.


  1. A. Mangal, E.A. Holm, A comparative study of feature selection methods for stress hotspot classification in materials, Integrating Mater. Manuf. Innovat. (7) (2018) 87–95,
  2. A. Mangal, E.A. Holm, Applied machine learning to predict stress hotspots I: face centered cubic materials, Int. J. Plast. (111) (2018) 122–134, 1016/j.ijplas.2018.07.013.
  3. A. Mangal, E.A. Holm, Applied machine learning to predict stress hotspots II: hexagonal close packed materials, Int. J. Plast. (2018),

Title: Laser Powder Bed Fusion Parameter Selection via Machine Learning Augmented Process Modeling

Presenter: Sandeep Srinivasan

Committee: Dr. Michael Groeber (advisor), Dr. Steve Niezgoda, Dr. Theodore T. Allen, Dr. Farhang Pourboghrat


  1. Purpose of the study: Laser Powder Bed Fusion (LPBF) Additive Manufacturing (AM) specifically for metals, is a highly active research area in the materials and manufacturing community, driven by promises of reduced lead time, increased design flexibility, and potentially location-specific process control. However, a complex processing space counters these benefits and results in difficulties when attempting to develop process parameter sets across different component geometries [1-4].
  2. Proposed approach: A procedure is developed for coupling physics-based process modeling with machine learning and optimization methods to accelerate searching the AM processing space for suitable printing parameter sets. The approach is demonstrated first on a few simple geometries that vary in size to show the methodology and then to a more complicated geometry to show the benefit of locally tailored process parameters on component processing history.
  3. Research method: The key steps in the scan parameter identification process developed in this work are (0) calibrate thermal model in parameter search range, (1) run thermal model for given set of scan parameters, (2) reduce the dimensions[5] of the thermal histories via Symbolic Aggregate Approximation (SAX)[6] and Principal Component Analysis (PCA)[7,8], (3) find Clusters in reduced local processing space using Density-Based Spatial Clustering of Applications with Noise (DBSCAN)[9], (4) compare to target clustering to obtain process similarity using a 5 N-component feature vector where N is the number of clusters,

(5) repeat 1-4 for M scan parameter sets, (6) fit statistical model to similarity metric data via Support Vector Regression (SVR)[10] and (7) search for optimal scan parameters using Gradient Descent (GD).

To demonstrate the utility of the proposed approach, local scan parameter modulation on a single 2D layer of an arbitrary geometry reflecting the complexities of typical AM components is exhibited. For simplicity, a component composed of rectilinear sub features to make tiling a simple procedure is designed. The complex geometry is first scanned with the reference scan parameters and then the complex geometry is tiled using square tiles of several sizes. The geometry is created in a way that these tiles of different sizes can be arranged such that they are printed in their entirety, which removes the issues associated with more complex scan parameter search. The tile with reference scan parameters is printed to obtain the target distribution of local processing states. After obtaining the optimized scan parameters, the tiles are positioned to create larger component geometry and are printed in sequence with their local scan parameters, effectively modulating the scan parameters across the component.

  1. Findings: (1) Locally modulated scan parameter strategy, the heterogeneity is significantly reduced with location (2) The speed associated with obtaining a similarity metric for a given scan parameter set is increased by proposing a regression approach (3) A threshold similarity, while minimizing the printing time, and similar processing conditions is obtained, which is the objective of the optimization (4) An additional positive aspect of the optimization is that the full component can be printed in less time with more uniform processing using this approach.
  2. Implications/Conclusions: (1) AM, particularly LPBF AM is a slow process and with the help of this technique, the parts can be optimized and manufactured quickly (2) LPBF of an aerospace alloy is sensitive and expensive and by using this proposed approach, robust parts can be manufactured in a cost-effective manner (3) Current work remains relatively simple in the geometries investigated and methods for comparing processing states, but more complicated geometries and metrics to produce uniform local processing are currently being investigated (4) The approach shows promising results in the ability to produce more uniform components in reduced printing time.


  1. T.J. Horn, O.L. Harrysson, Overview of current additive manufacturing technologies and selected applications, Science Progress 95 (3), 2012, 255-282.
  2. W.E. Frazier, Metal Additive Manufacturing: A Review, Journal of Materials Engineering and Performance 23 (6), 2014, 1917-1928.
  3. W.J. Sames, F.A. List, S. Pannala, R.R. Dehoff, S.S. Babu, The metallurgy and processing science of metal additive manufacturing, International Materials Reviews 61 (5), 2016, 315-360.
  4. D. Herzog, V. Seyda, E. Wycisk, C. Emmelmann, Additive manufacturing of metals, Acta Materialia 117, 2016, 371-392.
  5. L. van der Maaten, E. Postma, J. van den Herik, Dimensionality Reduction: A Comparative Review, 2009.
  6. J. Lin, E. Keogh, S. Lonardi, B. Chiu, A symbolic representation of time series, with implications for streaming algorithms, DMKD ’03 Proceedings of the 8th ACM SIGMOD Workshop on Research Issues in Data Mining and Knowledge Discovery, 2003, 2-11.
  7. K. Pearson, On lines and planes of closest fit to systems of points in space, Philosophical Magazine 2, 1901, 559-572.
  8. H. Hotelling, Analysis of a complex of statistical variables into principal components, Journal of Educational Psychology 24, 1933, 417-441.
  9. Ester, Martin; Kriegel, Hans-Peter; Sander, Jörg; Xu, Xiaowei (1996). Simoudis, Evangelos; Han, Jiawei; Fayyad, Usama M. (eds.). A density-based algorithm for discovering clusters in large spatial databases with noise. Proceedings of the Second International Conference on Knowledge Discovery and Data Mining (KDD-96). AAAI Press. pp. 226–231.
  10. Alex J. Smola and Bernhard Scholkopf, A tutorial on Support Vector Regression, Statistics and Computing 14: 199–222, 2004.
Category: Seminars