About A3SmartML

The overall objective of the project is to develop novel and generalising approaches for smart multidimensional measurements with significantly reduced experimental effort and validated uncertainty evaluation.

The specific objectives of the project are:

1. To develop automated machine learning approaches for the solution of general linear inverse problems based on digitised measurements, arising in scanning hyperspectral imaging or photocurrent mapping of semiconductors. To go beyond the state of the art by basing and extending these approaches with techniques from compressed sensing, low-rank reconstruction, and deep learning. To also train machine learning models to represent prior knowledge and statistically regularise the inverse problem, using both real and synthetic data.
2. To develop uncertainty evaluation methods for the approaches in objective 1 and create generalised tools to metrologically validate the uncertainty evaluation. These tools would include statistical uncertainty assessments and calibration analysis, as well as have the capacity to validate the robustness and reliability of the machine learning model.
3. To develop automated and adaptive measurement strategies for the methods developed in objectives 1 and 2. To achieve a targeted measurement uncertainty while satisfying the constraints of specific metrology applications, such as scanning hyperspectral imaging or photocurrent mapping of semiconductors. To also define a digitised measurement framework to include data management strategies and appropriate automated feedback procedures.
4. To implement the approaches developed in objectives 1 to 3, for metrological applications, such as scanning hyperspectral imaging and photocurrent mapping of semiconductors. To design automated and adaptive measurement strategies by taking into consideration any physical constraints imposed by measurement instruments or methodologies of the applications and develop experimental prototypes.
5. To facilitate the take up of the technology and measurement infrastructure developed in the project by standards developing organisations and end users (semiconductor industry, medical imaging etc.) and to disseminate good practice guidelines and a public database for reference data to all relevant stakeholders.

Solving high-dimensional linear inverse problems using ML for smart measurements

To achieve reductions in measurement times and costs for multidimensional measurements while maintaining their accuracy requires smart sub-sampling strategies and hence efficient data reconstruction and data completion tools. This involves the solution of high-dimensional, often ill-posed, linear inverse problems. Previous approaches include end-to-end solutions using trained neural networks, conventional regularisation approaches, or Bayesian inferences using standardised priors. The novel approaches to be developed in this project comprise the use of ML (i) to train a regulariser which is optimally adapted to the properties of the considered class of inverse problems, and (ii) to employ ML generative models such as Variational Auto Encoders, Generative Adversarial Networks, or Diffusion Models as priors in a Bayesian inference to represent a large database with physically meaningful solutions. In these ways, modern ML techniques are combined with established tools from compressed sensing, regularisation, and Bayesian statistics to further enhance these techniques and enable smart measurements.

Uncertainty evaluation and validation for smart measurements

Uncertainty evaluation for linear inverse problems with sub-sampled data is challenging since these problems are typically ill-posed. For the approaches developed in objective 1, reliable uncertainty evaluation depends on assumptions about the data and the uncertainty introduced through the employed ML technique. One challenge lies in the huge number of parameters, which prevents established tools from statistics (as well as of the GUM) to be immediately applied. For the novel Bayesian approach using a generative ML model as a prior, approximations in terms of Gaussian mixtures, and Monte Carlo sampling procedures will be developed, which approximately sample from the posterior. For the developed regularisation techniques, scalable uncertainty evaluation methods for deep neural networks such as Monte Carlo Dropout, Bayesian Neural Networks, or Deep Ensembles, and statistical interpretations of regularisers, such as Gibbs priors, are employed. Appropriate validation techniques are also explored that ensure the reliability of the estimated uncertainty.

Adaptive and uncertainty-aware measurement strategies

Adaptive measurement procedures will transform traditional scanning methods beyond the current state of the art by minimising data acquisition efforts in real time, significantly enhancing both efficiency and precision. Traditional scanning methods often involve uniform, time-intensive measurement processes. Techniques that account for spatially resolved uncertainty, however, enable the system to intelligently adjust scan parameters based on the specific features or regions. In measurements of complex materials, such as heterogeneous surfaces or intricate doping profiles, adaptive procedures will dynamically refine the resolution or measurement density, significantly shortening overall measurement times. This shortening allows to tackle more complex or dynamic systems with unprecedented speed and accuracy. The development of such smart, and possibly adaptive, measurement strategies is based on improved solutions of the corresponding inverse problems, along with a reliable uncertainty evaluation of the solutions.

Demonstrating the efficiency of smart measurements and their generalisability

Advancing applications such as scanning hyperspectral imaging for nanoscale material characterisation and photocurrent mapping of semiconductors by smart measurements will be demonstrated by generating experimental prototypes. This involves practical implementation of the developed data analysis and uncertainty evaluation approaches in the corresponding measurement devices and the novel development of automated processing techniques for data acquisition and analysis. Furthermore, the generalisability of the smart measurement techniques developed will be demonstrated by exploring their applicability for several additional use-cases, such as hyperspectral imaging scatterometry, laser induced breakdown spectroscopy, laser intensity modulation methods and X-ray scattering computed tomography, resulting in a generalisation pipeline that applies to a wide range of applications and technologies.

Outcomes for industrial and other user communities

The developments in this project will produce generalising approaches for more efficient and trustworthy measurements in multidimensional measurement scenarios that are relevant for widespread industrial application. The ML methods developed for the involved inverse problems, the corresponding uncertainty evaluation and the adaptive measurement strategies will be summarised and published in good-practice-guidelines. For nanoscale material characterisation, represented by scanning hyperspectral imaging, and the photocurrent mapping of semiconductors, experimental prototypes will generate immediate impact to manufacturers and end-users by demonstrating the efficiency of the implementation and improvements in measurement speed and resolution. Hyperspectral imaging scatterometry promises to be an extremely effective tool for quality control in the manufacture of Augmented Reality see-through optics. The prototypes, in conjunction with the good-practice guidelines and the generalisation pipeline produced within the project, facilitate the implementation of the developed approaches in user facilities, such as synchrotrons and storage rings, to increase the end-user throughput and improve the overall quality of the measurements conducted. Faster and more adaptive measurement processes will allow materials developing industries, such as novel semiconductor applications or power electronics, to perform rapid and reliable quality assessments on production lines, ensuring higher efficiency and minimising bottlenecks.

Outcomes for the metrology and scientific communities

The outcomes of this project will provide a common understanding among European NMIs/DIs on how to conduct trustworthy smart measurements by applying adaptive and uncertainty-aware measurements in multidimensional metrological applications. Moreover, the collaboration of European NMIs, industry and end-user facilities to develop experimental prototypes for scanning hyperspectral imaging and photocurrent mapping of semiconductors ensures a leading role of the European Union in these measurement capabilities and scientific developments. The produced good-practice guidelines can be easily adapted by the metrological and scientific communities. Research papers will be published in high impact peer-reviewed journals, and as part of the knowledge transfer, workshops on uncertainty evaluation and smart measurements will be organised and held, to which representatives of industry (both manufacturers and users), academia and NMIs will be invited. The provided open-source software and reference database will facilitate further method development of measurement capabilities in metrology through direct access to research results and reproducible methodologies. Results will be disseminated to the EMN MATHMET and as well as to the International Academy for Production Engineering (CIRP), which will make them accessible to a wider audience including stakeholders from all these networks. The collaboration of European NMIs in this project will increase their visibility and authority in drafting common regulations and guidelines. This will improve the competitiveness of the European economy, and it will lead to a more intense international cooperation.

Outcomes for relevant standards

The consortium will promote the results and outcomes of this project within the standardisation community and will provide input into the standardisation process. The participants of the project are active in the JCGM WG1, which has responsibility for the GUM and its supplements. These documents mark the de facto standard for uncertainty evaluation in metrology and are used worldwide at all levels of the measurement chain, from NMIs to industry. Furthermore, the results of the project will be disseminated to VDI/VDE, ISO and CEN working groups. For ISO, the relevant standards that are in preparation/revision will be identified, and the results of this project relevant for these standards will be proposed to the appropriate working groups or committees. The participants will also present the outputs of the project in CIRP, IMEKO, EURAMET and other networks, where they are active. All these activities will ensure the uptake of the project’s outcomes.

Longer-term economic, social and environmental impacts

The improved capabilities at NMIs and DIs, which will be provided by this project, will strengthen the competitiveness of Europe’s industry and enable it to reduce measurement times and costs for multidimensional measurement scenarios and hence support the EU goal to become climate neutral by 2050. The developments will also lead to a reduction in production time and a reduction in costs for many quality control and defect detection applications in industry. Therefore, novel applications and systems in several sectors including medical, optical and precision instruments, as well as electronics, telecommunication and clean energy will emerge. Improvements in the reliability, efficiency and speed of production processes will also significantly decrease the scrap rate and reduce the energy needed for production. Positive social effects will result from the impact of high-end optical and semiconductor components for smartphones, wearables, artificial reality glasses, healthcare instrumentation and from the production of new photovoltaic systems, leading, in the longer term, to better healthcare, more affordable and accessible clean energy and the creation of highly skilled jobs in Europe.