Knowledge based data boosting exposition on CNT-engineered carbon composites for machine learning
- PDF / 1,193,103 Bytes
- 11 Pages / 595.276 x 790.866 pts Page_size
- 55 Downloads / 152 Views
ORIGINAL RESEARCH
Knowledge based data boosting exposition on CNT-engineered carbon composites for machine learning Sunil Chandrakant Joshi 1 Received: 19 May 2020 / Revised: 22 July 2020 / Accepted: 7 August 2020 # Springer Nature Switzerland AG 2020
Abstract Machine learning (ML) is useful in predictive analytic or prognostic modeling for materials and engineering. It is, however, challenging to gather sufficient and representative data. Experiments are possible only in small numbers due to specialty materials, manufacturing, infrastructure, and testing involved. Simulation and numerical models need skills and appropriate validation. If the dataset at hand is too small in size to train ML, professionals tend to create synthetic data, which may not necessarily meet the quality required of the new data. A Knowledge-based Data Boosting (KDB) process, named COMPOSITES, that rationally addresses data sparsity without losing data quality is systematically discussed in this paper. A study on inter-ply fracture toughness of carbon nanotube (CNT)-engineered carbon fiber reinforced polymer (CFRP) composite laminates is used to demonstrate the KDB process. This involved strengthening of inter-ply interfaces using CNT advocated for improving delamination resistance of the CFRP composites. It is demonstrated that the KDB process helped augment the dataset reliably and improved the best fit regression lines. The process also made it possible to define boundaries and limitations of the augmented dataset. Such sanitized dataset is certainly valuable for prognostic modeling. Keywords Carbon composites . Knowledge-based data boosting . Prognostic modeling . Machine learning . Fracture toughness
1 Introduction The field of machine learning (ML) is fairly well developed where a machine is allowed to learn from examples and history without being explicitly programmed. ML is influencing many application areas in engineering, including composite materials [1], materials design [2], manufacturing [3, 4], optimization [5], and prognosis [6–9]. ML depends heavily on algorithms and data. Many algorithms and programming frameworks are now available (e.g., [10]). However, it is impossible for a machine to learn without data. Such dataset/s shall be accurate and sufficient for any sound ML exercise. ML requires to be built and progressed through 3 stages, namely training, validation, and testing. All these 3 stages require data. Their proportion, in general [11], is shown in
* Sunil Chandrakant Joshi [email protected] 1
School of Mechanical and Aerospace Engineering, Nanyang Technological University, Singapore, Nanyang Avenue, Singapore 639798, Singapore
Fig. 1. It is clear that the majority of the data is needed for training ML models. In many cases, synthetic data [12] or a data augmentation process [13] is adopted. Use of synthetic data, fully or partially, is not advisable for engineering prognosis applications, as the risk involved is high. Data augmentation certainly works well with imaging. Popular augmentation techniques, such
Data Loading...
