MecaNano Pitching Event + Round Table on Machine Learning Applications (WG1–WG3)

Friday 12 Dec 2025, 10:00 → 15:15 Europe/Vienna

Online

Claus Trost (Erich Schmid Institute of Materials Science of theAustrian Academy of Sciences), Edoardo Rossi (Università degli Studi Roma Tre)

This online event is designed to foster collaboration across WG1, WG2, and WG3 by connecting researchers with varying levels of expertise in machine learning. Participants will have the opportunity to present their problems, datasets, or methodological questions in short 5-minute pitches.

The session will be guided by machine learning experts, who will provide feedback and facilitate discussion during a structured round table. The goal is to create a space where non-experts can receive targeted advice, experts can share best practices, and all participants can identify opportunities for cross-disciplinary collaboration.

The event will combine concise pitch presentations with an interactive round table discussion, ensuring that both practical challenges and conceptual developments are addressed.

1 Welcome & Objectives

A short introduction to the event, outlining the goals of the pitching session and the expected outcomes for participants across WG1–WG3.

Speakers: Claus Trost (Erich Schmid Institute of Materials Science of theAustrian Academy of Sciences), Prof. Edoardo Rossi (Università degli Studi Roma Tre)

2 Machine Learning Strategies for Micromechanical Phase Classification from Nanoindentation Data

This work investigates how machine learning can be used to classify mechanical phases in WC–Co hardmetals from nanoindentation data. High-speed nanoindentation provides depth-resolved mechanical responses that contain rich information about local microstructural behavior. By focusing on the full hardness–depth curves instead of single-point values, meaningful features can be extracted for phase identification.

The study combines unsupervised and supervised learning to explore different classification strategies. Data cleaning and clustering are first used to identify representative mechanical behaviors and label them as WC or binder. Convolutional neural networks trained on curve-derived images provide phase classification with confidence scores, enabling the detection of uncertain or transition regions near interfaces. A PCA–Random Forest approach offers a complementary and faster alternative with comparable performance.

The main open questions concern model generalization, overfitting, and data dependence on experimental setup.

Speaker: Laia Ortiz Membrado (Universitat Politècnica de Catalunya)

3 Synthetic Nanomechanical Data and Kernel-Averaged Mechanical Mismatch for Robust Phase Clustering

We present an unsupervised machine learning framework to enhance phase identification in nanoindentation mapping of multiphase materials, combining physics-informed features with synthetic data generation. Central to this approach is the Kernel-Averaged Mechanical Mismatch (KAMM), a spatial descriptor that quantifies local gradients in elastic modulus (E) and hardness (H). KAMM is incorporated into multiple clustering algorithms—including k-means, Gaussian Mixture Models, DBSCAN, and agglomerative clustering—to evaluate its effect on clustering robustness and accuracy. To systematically benchmark performance, we generate synthetic indentation datasets with controlled gradients in mechanical properties, emulating multiphase composites and interfacial zones. These datasets enable parametric studies of clustering behavior under varying phase contrast, noise, and sampling density. Results show that incorporating KAMM consistently enhances intra-cluster compactness and inter-cluster separability, particularly in low-contrast or graded materials, enabling a more quantitative benchmarking of unsupervised learning approaches for nanomechanical data analysis. Finally, we demonstrate a lightweight, interactive application that integrates these clustering workflows—allowing users to visualize indentation maps, compare algorithms, and assess the impact of KAMM in real time for advanced phase-mapping analysis.

Speaker: David MERCIER (Ansys part of Synopsys)

4 Correlated nanoindentation mapping and microstructural chracterization of advanced ferritic-austenitic stainless steels.

Understanding the origins of local mechanical behaviour is crucial for developing novel steel grades. Here, the mechanical properties of metastable, lean alloyed ferritic-austenitic stainless steels were investigated using nanoindentation mapping. The level of metastability was modified through alloying, resulting in different austenite fractions ranging from 17 to 50 %. For nanoindentation, A 50 x 40 indentation grid (2000 indents) was performed on each sample. The analysis combined nanoindentation mapping, electron microscopy, and K-means clustering to extract meaningful insights from the extensive dataset. Electron backscatter diffraction (EBSD) provided detailed microstructural information, such as phase distribution and crystallographic orientation, while nanoindentation mapping enabled the phase-specific analysis of mechanical properties. Additionally, variations within the ferrite grains were detected with the steel containing low austenite phase fractions, however these observed varitions within the ferrite grains remain unclear.

Speaker: Rahul Cherukuri (Universität Kassel, Tampere University)

5 Integrating CSM Nanoindentation and Materials Databases for Machine Learning-Based Phase Insights

A machine learning-assisted approach for phase identification in materials could make use of nanoindentation mapping, particularly continuous stiffness measurement (CSM) data. The richer set of mechanical features accessible through CSM such as hardness, modulus, indentation depth, contact stiffness, energy dissipation, and the evolving H/E ratio combined with freely available materials databases, might allow likely phases in a sample to be suggested. Nanoindentation points could be clustered, providing probabilistic associations with known phases and an interpretable confidence estimate. Optional guidance could be given by specifying the expected number of phase groups or approximate ranges of mechanical properties, while the method could also operate without prior assumptions. As a potential extension, learning-based checks could be included to identify common indentation artifacts such as pile-up or sink-in. This conceptual vision illustrates how the combination of rich mechanical mapping with external material knowledge could support exploratory phase identification in a flexible and informative manner.

Speaker: Tizian Arold

11:15 Coffee break

6 Analysis of deformation mechanisms in Ni-based cemented carbides by means of micropillar compression and Machine Learning techniques

Cemented carbides are multiphase materials constituted by a hard ceramic phase,
commonly tungsten carbide (WC), embedded in a metallic matrix, being the mostly
used Co. The mechanical properties achieved from these two constitutive phases,
including an excellent hardness, strength, toughness, wear and abrasive resistance,
makes the WC a highly demanded material for industrial applications, such as cutting
tools, molds, mining parts or industrial nozzles, among others. However, cobalt is
considered as a Critical Raw Material (CRM) and carcinogenic-mutagenic and toxic to
reproduction material (CMR), hence new alternatives are emerging, being the most
promising one the nickel due to its similarity – in terms of mechanical, electrical and
chemical properties – and an improved corrosion resistance. Despite this situation,
cemented carbides with Ni-based metallic matrix remain unstudied, including its
mechanical behavior under stress, strongly influenced by the surrounding deformation
mechanisms between its constitutive phases.
The micropillar compression technique enables the application of a controlled stress
to a small volume of material, allowing the deformation mechanisms to be reproduced
and analyzed through the stress-strain curves. Moreover, when using an in-situ
nanoindenter, the live SEM monitoring of the micropillars during the experiment allows
the accurate correlation of the pop-in events, visible in the stress-strain curves, with
the different deformation mechanisms. Establishing the proposed correlation
becomes particularly relevant when examining its dependence on other variables,
such as testing temperature. Since different deformation mechanisms are expected to
operate under varying thermal conditions, this correlation offers a valuable means of
distinguishing between them.
As all the aforementioned procedures can be time-consuming, Machine Learning is
expected to improve the efficiency of the analysis by 1) facilitating the identification of
pop-in phenomena in the stress-strain curves, 2) facilitating the identification of
deformation mechanisms in the live SEM monitoring and 3) automating the procedure
of correlating pop-ins and deformation mechanisms. An additional area for
improvement with Machine Learning techniques could be the cross-section analysis by
means of Deep Learning, to automate the identification of deformation mechanisms
through cross-section images of the post-mortem micropillars.

Speakers: Emilio Jiménez-Piqué (Universitat Politècnica de Catalunya), Francesc Barberá-Flichi (CIEFMA, Department of Materials Science and Engineering, Universitat Politècnica de Catalunya - BarcelonaTech, Campus Diagonal Besòs-EEBE, 08019 Barcelona, Spain 2Barcelona Research Center in Multiscale Science and Engineering, Universitat Politècnica de Catalunya, Campus Diagonal Besòs − EEBE, Barcelona, 08019, Spain), Laia Ortiz Membrado (Universitat Politècnica de Catalunya), Luis Llanes (1CIEFMA, Department of Materials Science and Engineering, Universitat Politècnica de Catalunya - BarcelonaTech, Campus Diagonal Besòs-EEBE, 08019 Barcelona, Spain 2Barcelona Research Center in Multiscale Science and Engineering, Universitat Politècnica de Catalunya, Campus Diagonal Besòs − EEBE, Barcelona, 08019, Spain)

7 Exploration of the hyperspace of refractory high-entropy alloys using high-throughput materials engineering and machine learning

There is a growing demand for metallic alloys with high strength and corrosion resistance at elevated temperatures for applications in aerospace, power generation, and chemical industries. With the ad-vancement of modern technologies, the need for metallic materials capable of operating at higher tem-peratures than current Ni-based superalloys continues to increase. Refractory high-entropy alloys (RHE-As), which have emerged over the past decade, represent a highly promising class of materials for ex-treme environments due to their high melting points and exceptional high-temperature strength.
In this study, we present a systematic exploration of the hyperspace of RHEAs within the Cr–Mo–Nb–Ta–V–W system. A materials library (ML) was fabricated using physical vapor deposition (PVD) on a silicon wafer, designed to produce continuous compositional gradients of each element in the range of 30–45 at.%. Assuming a compositional resolution of 1 at. %, this corresponds to approximately 35,000 distinct alloys. Synthesizing the same number of alloys via conventional methods such as arc melting or powder metallurgy, at a rate of one alloy per day, would take roughly 136 years. The co-sputtering process was calibrated to achieve an equimolar composition and maximum configurational entropy at the center of the ML (Fig. 1).
In the first stage, selected regions of the ML were characterized by X-ray fluorescence (XRF) to determine their chemical composition. Structural characterization was performed using X-ray diffraction (XRD), and the resulting diffractograms were analyzed via Le Bail fitting - a full-profile method allowing extraction of crystallographic parameters such as lattice constants and crystallite size. Subsequently, high-throughput nanoindentation was employed to assess mechanical properties. The resulting materials dataset was then used to train an artificial neural network (ANN) to predict mechanical properties of alloys extending be-yond the compositional space directly explored in the ML.

Speaker: Krzysztof Wieczerzak (Department of Materials Science, Faculty of Mechanical Engineering and Aeronautics, Rzeszow Universi-ty of Technology, al. Powstancow Warszawy 12, 35-959 Rzeszow, Poland)

8 Exploring the extrapolation limits of machine learning models for refractory complex concentrated alloys using high-throughput experiments

Refractory complex concentrated alloys (RCCAs) are emerging as key candidates for structural applications under extreme conditions, where both strength and stability beyond 1000 °C are required. Yet, the immense compositional space of these alloys (spanning billions of possible combinations) makes their systematic exploration by conventional methods infeasible. High-throughput materials science offers a solution: combinatorial PVD synthesis enables the fabrication of thin-film materials libraries with continuous composition gradients, effectively covering binary or ternary systems within a single deposition. However, when moving to higher-order systems, the composition gradients attainable on planar substrates represent only a two-dimensional projection of a much higher-dimensional space.

In this study, we focus on a selected quaternary refractory system (A–B–C–D) and synthesize a series of combinatorial thin-film libraries covering different compositional regions. Using high-throughput mechanical characterization, we will build a large dataset for training machine learning models that predict mechanical properties across this multicomponent system. The main objective is to determine how far one can extrapolate meaningfully from a given training data cloud, quantified using the Mahalanobis distance as a metric of compositional similarity. Moreover, we will assess whether combining multiple data clouds from distinct composition regions improves the extrapolation capability of the models.

This approach will provide new insights into the limits of machine learning prediction in complex alloy systems and guide the efficient design of high-dimensional RCCAs through data-driven methods.

Speaker: Hanna Szebesczyk

9 Accelerating the discovery of ductile refractory complex concentrated alloys via high-throughput experiments and machine learning

Refractory complex concentrated alloys (RCCAs) represent a promising class of materials for next-generation high-temperature applications, where conventional Ni-based superalloys and ceramics reach their performance limits. However, their development is hindered by intrinsic brittleness at low temperatures and the vastness of their compositional design space, which exceeds billions of possible combinations.
To address these challenges, we propose a high-throughput experimental and data-driven approach to explore the interplay between chemical composition, microstructure, and ductility in model binary refractory systems. Thin-film materials libraries will be synthesized by combinatorial PVD, enabling systematic mapping of structure–property relationships across wide composition gradients. The rate sensitivity of deformation, which correlates with ductility, will be rapidly quantified using nanoindentation-based high-throughput mechanical screening.
The resulting datasets will serve as a foundation for training machine learning models capable of predicting mechanical properties not only within a single binary system but also across different ones. We aim to evaluate (i) how transferable such models are between systems with shared elements (e.g., from A–B to A–C or B–C), and (ii) whether integrating data from multiple binary systems enables reliable extrapolation to ternary or higher-order alloys.
Ultimately, this work seeks to establish a generalizable, scalable workflow that integrates combinatorial synthesis, rapid mechanical characterization, and machine learning to accelerate the discovery of ductile RCCAs for extreme-temperature environments.

Speaker: Maria Kanczewska

12:30 Lunch Break

10 Flue-gas desulfurization

Scrubbing with seawater is a reliable technology for flue-gas desulfurization in power plants or in marine engines. Then, the removed sulphur is released to the atmosphere, or effluents are released to the sea. In both cases, stringent legal limits for both emissions and effluents are in effect that must be met. There are many input parameters (sulfur content, flow rate, temperature of the flue gas), processing parameters (seawater flow rate in scrubber, temperature, height of scrubber), and environmental parameters (air temperature, air velocity, humidity for emissions, and seawater temperature, current velocity, salinity, and pH for effluent dispersion). We would like to use ML to provide us with the proper processing parameters, given the input and environmental parameters, that obey the stringent legal limits.

Speaker: Pavlos Stephanou

11 Bacterial Cellulose Polymers for PEM Fuel Cells

The utilization of cellulose-producing bacteria in decarbonizing hydrogen technologies, particularly for proton exchange membranes (PEMs) in fuel cells and electrolyzers, is an alternative for more sustainable hydrogen production. Numerous research institutions and universities have investigated the cultivation of these non-pathogenic, naturally occurring bacteria as eco-friendly substitute. Key bacteria have been extensively studied for their ability to produce bacterial cellulose (BC) with high purity, mechanical strength, and biocompatibility. These properties can be further optimized through
co-culturing with complementary microorganisms, potentially improving membrane performance metrics like proton conductivity and thermal stability.

Membrane fabrication will employ electrospinning to achieve precise control over nanofiber morphology, influenced by parameters such as solution viscosity, solvent and additive selection, applied voltage, flow rate, and ambient conditions like humidity and temperature. This method enables the creation of high surface area, tunable structures suitable for PEM applications.

To advance this study, a comprehensive database will be constructed, encompassing cultivation variables ( pH, temperature, humidity) alongside electrospinning and post-processing data (fiber diameter, porosity, mechanical strength, biodegradation rate, among others). This repository will facilitate the analysis, including machine learning models for parameter optimization, predictive modelling of membrane performance, and identification of key influencing factors.

Speaker: Lidia Valdés Llorente (Technical University of Leoben)

12 Machine Learning-Assisted Wear Behavior Evaluation of P/M Ti-Matrix Hybrid Reinforced Composites

Titanium is widely used as a matrix material in composites because of its favorable mechanical, physical and chemical properties, yet its cost and limited wear resistance and high-temperature performance restrict some applications. Ceramic-reinforced titanium matrix composites offer a practical route to improve wear behavior.
This study focuses on powder-metallurgy titanium-matrix composites with varying amounts of Al₂O₃, CeO₂ and few-layer graphene, and on how composition and grain size affect wear under 1–3 N loads and 10 m / 50 m sliding. Using the available lab measurements (volume loss, density and hardness), we will compute physics-based summaries such as the Archard wear coefficient and build conservative, interpretable predictive models. Given the modest dataset, we will treat machine learning as a decision-support tool: Gaussian Process Regression for uncertainty-aware prediction, tree-ensemble methods for ranking influential factors, and simple linear or regularized models as baselines. Models will be cross-validated and interpreted with partial-dependence and Shapley-value analyses. In this phase microscopy is not available, so recommendations will rely on measurable metrics and model interpretation. Surrogate-model suggestions will guide a short list of compositions for targeted experimental validation. All analysis will be performed in MATLAB and reported with quantified uncertainty.

Speaker: Deniz Uzunsoy (Bursa Teknik Üniversitesi)

13 Machine Learning vs. Constitutive Models: Can Data Replace Equations in Finite Element Analysis?

Recent advances in machine learning (ML) have created new opportunities for modeling the mechanical behavior of polymers and composites. Conventional constitutive models require assumptions, parameter calibration, and often struggle with nonlinear, rate-dependent, and temperature-sensitive responses. ML approaches, trained directly on experimental data, offer an alternative, but important questions remain:
Can ML reliably replace constitutive equations, and how can these models be integrated into finite element analysis?
What level of physical constraints, data quality, and multi-axial representation is necessary to ensure accurate and stable predictions?
In this round-table discussion, I will outline current practices in ML-based material modeling, highlight potential advantages and limitations, and invite debate on whether data-driven constitutive laws can evolve into robust, physics-guided tools for FEM in real engineering applications.

Speaker: Özgen Çolak Çakır (Istanbul Technical University)

14 Machine Learning for Adaptive Behavior in Smart Soft Materials

We are developing a new class of soft, adaptive composite materials with tunable stiffness and embedded electromechanical responses. Under controlled mechanical activation, these materials generate complex multimodal signals that are strongly coupled to their internal structure.
Before any robotic finger or human testing, our primary focus is AI for Materials: using machine learning to understand, map, and predict the behaviour of nonlinear soft composites. Building on multiscale characterisation data (electron microscopy and in-situ/mechanical testing), we aim to explore modelling approaches capable of:
• identifying optimal and smooth modulus gradients,
• analysing particle distribution and agglomeration patterns,
• learning structure–property–response relations under different activation states,
• detecting emergent nonlinear behaviour in adaptive composites.
While the core of this work is material-centred, the resulting models may also support downstream AI-assisted control or interpretation frameworks in soft robotic applications.
We welcome collaboration from researchers working on:
• physics-informed ML,
• data-efficient modelling,
• multimodal data fusion,
• AI for materials and adaptive systems.

Speaker: Dr Ayse Cagil Kandemir (TED University)

Discussion and Reflection

Conveners: Claus Trost (Erich Schmid Institute of Materials Science of theAustrian Academy of Sciences), Prof. Edoardo Rossi (Università degli Studi Roma Tre)