Knowledge-driven AI

Background / Motivation

The past decade has seen tremendous advances in machine learning (ML) and artificial intelligence (AI), especially in the application of deep learning to natural language processing and computer vision. This data revolution has led to massive productivity in ways such as accelerated materials development cycles, improved predictive maintenance in industries such as aerospace, and enhanced process optimisation in advance manufacturing.  

However, perennial concerns remain over some fundamental limitations in conventional ML: (i) the requirement for large and typically expensive-to-acquire datasets; and (ii) a lack of generalisability and interpretability. A potential solution to alleviate these limitations is to incorporate knowledge with AI. Such knowledge-driven AI (K-AI) can comprise both scientific and human-centric knowledge. 

IHPC, with its expertise in engineering and numerical theory-based simulation, cognitive and social sciences, and machine learning, is in a unique position to combine deep domain knowledge in areas such as physics, chemistry and social sciences with novel AI models and methods to create greater impact across different domains.


IHPC’s K-AI efforts (Fig 1) include: 

Knowledge-Guided Forward Modelling - Integrating domain knowledge with AI for forward prediction

  • Hybridisation of AI and Simulation
  • Knowledge-guided design of AI model architecture
  • Knowledge-constrained training of AI models

Knowledge-Guided Inverse Design - Utilising domain knowledge during optimisation to achieve optimal design solutions and/or to accelerate the design process

  • Knowledge-guided design of algorithms for more optimal search space exploration 
  • Knowledge-guided reduction of search space for faster optimisation

Knowledge-Guided Representation Learning - Utilising domain knowledge to better acquire and represent information for incorporation into the AI models for both forward modelling and inverse design

  • Knowledge-guided data representation for improved feature construction and selection for AI models 
  • Strategies for knowledge acquisition and extraction from domain experts or indirect sources of information (e.g., reasoning for general expertise)

Fig 1. IHPC's Research Foci within K-AI


Knowledge-driven AI

IHPC has been promoting K-AI through multi-party collaborations across departments within IHPC, across A*STAR’s Research Institutes, and external partners. While a majority are in the engineering domain like materials development, advanced manufacturing and aerospace, the K-AI collaborations include other domains such as healthcare in partnership with local clinical institutes. 

In the following section, using advanced manufacturing as an example, we illustrate how K-AI can be used for (i) parts design and process optimisation; (ii) engineering process control; and (iii) defect detection post-fabrication.
  • Industrial Digital Design and Additive Manufacturing Workflow

Components of AM workflow
Fig 2. Example of components of AM workflow that can be modelled with a K-AI 
surrogate model for faster and more cost-effective part design

In order to industrialise additive manufacturing, it is necessary to develop intelligent software tools that allow industry to exploit the high level of control that additive manufacturing processes grant over both geometry and materials while incorporating manufacturability considerations upfront during the design stage (Fig 2). By working with the A*STAR's Institute of Materials Research and Engineering (IMRE) and Advanced Remanufacturing and Technology Centre (ARTC), and other Institute of Higher Learning (IHL) partners like the National University of Singapore (NUS) and Singapore University of Technology and Design (SUTD), physics-based domain knowledge and AI methods are used to construct surrogate models of additive manufacturing processes within a typical additive manufacturing workflow. This is a proof-of-concept demonstration of the possibility for fast and cost-effective design for additive manufacturing processes such as Selective Laser Melting (SLM).
  • ML-assisted Control of Shot Peening Process

Model-Predictive Control Framework
Fig 3. Model-Predictive Control framework for utilising K-AI models 
to enhance engineering process control

Advanced manufacturing processes such as shot peening are highly complex and can be difficult to deploy in the industry. For example, shot peening relies heavily on post processing quantification such as Almen strip testing and saturation curve development which are time consuming and costly. Real-time monitoring and control for advanced manufacturing processes such as shot peening can also reduce the costs associated with characterisation and re-work. 

In our collaboration with ARTC and industry partners, a model predictive control framework (Fig 3) was developed to produce a desired output intensity. This was achieved by utilising a physics-based AI model and real-time sensor data to control process parameters optimally. This is a demonstration of the use of physics-based AI models for engineering process control, with potential for expansion to other engineering processes.

  • Automated Defect Detection with X-Ray CT Images

K-AI for Defect Identification
Fig 4. Example of utilising K-AI for defect identification which is a common 
inverse modelling problem for non-destructive testing (NDT)

Defects in components fabricated by additive manufacturing are a common issue, and these defects can significantly impact the components’ quality and reliability (Fig 4). X-Ray CT scans are a way to detect these defects in the components prior to further processing for reduced cost and improved efficiency. By utilising physics-based simulation models, additional simulated data and parameters can be generated, which will help in building better AI models for the inverse problem of automated defect detection from X-Ray CT images. Further domain information such as symmetry and invariances can be incorporated into such AI models to improve their generalisability.

Quantum Computing

Quantum computers have the potential to produce tremendous improvements in computational performance for a wide range of problems in various fields, including optimisation, machine learning, chemistry, finance, and healthcare. As the performance of quantum hardware improves, so does the possibility of implementing quantum algorithms that outperform their classical counterparts. There is currently an ongoing race amongst various academic research groups, companies and national research labs to produce the first demonstration of a practical quantum advantage – a milestone in which a quantum computer solves a problem of practical interest with a performance that cannot be matched by classical computers. Such a demonstration would involve: (i) using actual quantum hardware to run such a quantum algorithm, and (ii) showing that classical computers cannot perform just as well.

IHPC’s Quantum Computing applications (Fig 5) include:

  • Simulations
  • Optimisation
  • Software for quantum control and readout
  • Horizontal effort to understand the power and limitations of near-term quantum computers
Benchmarking against classical algorithms and high-performance computing (HPC) systems is an integral part of these application areas.

Quantum Computing Applications Areas
Fig 5. IHPC's focus areas for quantum computing

Collaboration Opportunities 

IHPC welcomes interested parties to collaborate in research and development relating to Artificial Intelligence and Quantum Computing, especially in identifying new technology areas to develop new applications, multi-disciplinary projects or commercialisation outcomes. Leverage our domain expertise and K-AI framework to solve interesting problem statements.  

For more info or collaboration opportunities, please write to