By Prof. Pasquale Franciosa, WMG, University of Warwick, UK

Summary

Prof. Pasquale Franciosa’s lecture explains how laser welding can be combined with machine learning, physical simulation, and sensing to support digital manufacturing. The emphasis is on metal joining for electric mobility, particularly battery packs and electric-drive components. The talk moves from laser-welding fundamentals to monitoring and predictive modeling, then presents two industrially oriented case studies.

1. Research context and scope

Franciosa describes his work at WMG, University of Warwick, as spanning fundamental laser–material research through applied industrial development and prototype validation.

The research covers:

• Laser–material interaction and the effects of process parameters and heat input.
• Advanced modeling and digital twins.
• In-process monitoring and control.
• Full-scale prototypes and production readiness.
• Testing and characterization, from material coupons to components.

The work is framed around sustainable mobility. Sustainability includes environmental impact, economic considerations, and social factors.

2. Laser-welding fundamentals and challenges

In laser welding, the laser acts as the heat source. Different materials absorb and reflect laser energy differently, so the same laser settings can produce different outcomes in steel, aluminium, or copper.

The lecture distinguishes three welding modes:

• Conduction mode: The material melts, but the process does not form a deep vapor cavity.
• Transition mode: The process approaches the material’s vaporization temperature.
• Keyhole mode: The material vaporizes and forms a narrow cavity, enabling deep penetration.

The presenter notes that single-pass laser welding can achieve very deep penetration; he cites a reported depth of about 70 mm in mild steel.

Keyhole welding is useful but difficult to control. The process involves solid material, molten metal, vapor, and sometimes plasma, with heat transfer, fluid flow, surface tension, and buoyancy all influencing stability.

An unstable keyhole can collapse and leave defects such as pores, inclusions, and voids. These can make a weld unacceptable under quality requirements. The lecture stresses that defect causes may vary widely, including gaps, incorrect parameters, mechanical variation, surface contamination, material changes, and thickness variation.

3. Laser sources, delivery, and beam shaping

A laser-welding system generally includes a laser source, an optical fiber, and beam-delivery optics. The talk contrasts older CO₂ systems with more common modern diode- and fiber-based sources. Wavelength matters because it affects how much energy a material absorbs.

Copper and aluminium are challenging with infrared lasers: the presenter cites absorption in the approximate 5–10% range. Blue lasers, around 450 nm, can be more effective for these materials, but he notes cost and system-efficiency limitations.

Laser systems can use different delivery arrangements, including:

• Fixed optics.
• Oscillating or “wobble” optics, which move the beam rapidly.
• Filler-wire delivery, used, for example, to adjust the molten pool’s composition when welding crack-sensitive aluminium.
• Remote scanning optics, which can move the beam rapidly over a larger work area.

High-power lasers require strict safety controls. The presenter describes industrial systems as typically enclosed and notes that handheld systems also raise safety considerations.

Beam shaping is presented as an important way to influence heat distribution and weld quality. Instead of a standard round, Gaussian-like spot, systems can create core-and-ring, donut, tailed, or other profiles. Dynamic beam shaping can change the spot shape during welding.

The intended benefits include improving weld shape, reducing porosity, stabilizing the keyhole, and potentially influencing microstructure and phase formation. The appropriate profile depends on the material and application.

4. Monitoring, machine learning, and digital manufacturing

Monitoring is essential because weld quality can be affected by many different process and material variations. Sensors provide data that can be used to identify defects or adjust the process.

Sensor approaches discussed include:

• Photodiodes: Low-cost devices that measure emitted light and provide electrical signals for analysis. The presenter shows examples where signal changes correlate with weld condition and gap size.
• Optical coherence tomography (OCT): Uses a low-power beam and interferometry to estimate keyhole depth. Measurements can be challenging because the keyhole moves and fluctuates.
• Vision systems and cameras: Can monitor welds and detect component positions, such as hairpins, before the laser fires. Correct positioning is particularly important for electrical components.
• Laser triangulation and image processing: Can provide geometric information that may be linked to weld quality.
• Thermal cameras and high-speed cameras: Thermal cameras can monitor temperature over a wide area, but readings are sensitive to surface emissivity. High-speed imaging can reveal process behavior, including during solidification.

The talk defines digital manufacturing in terms of connecting data, computation, and the physical process to make production more flexible and resilient, while reducing defects and waste and controlling costs.

Franciosa highlights a concise framework: “the intelligent connection of perception to action.” In this framing:

• Intelligence comes from AI and physical models.
• Connectivity links the physical process to digital tools through sensing.
• Perception means capturing and interpreting process conditions.
• Action means changing settings such as laser power, speed, or frequency.

A major proposal is a hybrid approach that combines data-driven machine learning with physics-based models. Machine learning can be fast after training, but it depends on having representative, good-quality data. Physical models can provide process understanding but may take hours or days to compute.

Important implementation challenges include sensor accuracy and speed, sufficient training data, model accuracy, computational latency, safety, ethics, and protection of industrial intellectual property.

The presenter argues that welding control must be fast enough to respond within the process’s short solidification times, which he puts at roughly a few milliseconds to around 10 milliseconds.

5. Modeling and machine-learning workflow

The lecture describes a general workflow for developing and deploying a monitoring or prediction system:

1. Define the process and quality problem. Identify the weld characteristics or defect classes that matter.
2. Run experiments under varied conditions. Change relevant inputs such as power, gap, speed, or beam profile.
3. Collect sensor and process data. Use appropriate monitoring equipment and check whether it captures the behavior of interest.
4. Extract features and label outcomes. Link the measured signals to known weld conditions or quality classes.
5. Train and validate models. Test multiple machine-learning methods rather than assuming one algorithm will work for every application.
6. Assess errors and data coverage. Overlapping classes create ambiguous “gray areas” that can lead to false positives, false negatives, and misclassification.
7. Improve the data when needed. Data augmentation can create additional examples from existing data, though it does not remove the need for representative real-world measurements.
8. Deploy and update. A trained model may be deployed as-is or continually updated as new data arrive.

In one battery-welding classification example, around 200 experiments were used to distinguish good welds, insufficient connection, and over-penetration. Several algorithms produced accuracies reported between approximately 89% and 97%, but none reached the presenter’s aspirational target of 100%.

The example illustrates that performance depends on the application, the quantity and quality of data, and how clearly the classes are separated.

The presenter describes reduced-order modeling as a way to make physical simulations more computationally practical. A reduced-order model can retain process-specific physics while reducing the number of equations that must be solved.

• Intrusive methods modify the underlying physical equations and can achieve high accuracy, but require substantial mathematical and implementation expertise.
• Non-intrusive methods use data generated by physical simulations and apply machine-learning techniques to approximate their results. In the example discussed, they were faster but less accurate than the intrusive approach.

Overall, the goal is to balance accuracy, computational cost, and the speed required for production use.

6. Case studies

Aluminium battery casing

The first case concerns crack-sensitive, recycled 6000-series aluminium used in battery-casing applications. The challenges include avoiding weld cracks and limiting thermal distortion while achieving a sound joint.

The research explored different welding strategies, including power modulation and beam oscillation. Oscillation can help achieve sufficient melting at higher travel speeds while managing heat input.

The presenter emphasizes that the final process was developed through a combination of experiments, parameter optimization, geometry selection, physical simulation, and data analysis.

Copper hairpin welding for electric drives

The second case concerns copper hairpins, which are challenging because copper reflects infrared laser energy and because component alignment and gaps can vary.

High-speed videos compare two beam profiles. One produced more spatter and a less stable molten pool; a different, simulation-informed profile produced a more stable process and better bridging across the gap.

The presenter reports that the improved beam profile reduced porosity and substantially increased joint strength in the tested comparison. The case demonstrates how process knowledge, imaging, simulation, and beam shaping can be combined to address real manufacturing variation.

7. Sustainability perspective

The lecture cautions against treating electric vehicles as impact-free simply because they have no tailpipe emissions during use. Manufacturing, raw-material extraction, and end-of-life disposal or recycling also contribute to environmental impact.

Laser processing is presented as important in electric-vehicle manufacturing, so improving its efficiency and weld quality may contribute to better production outcomes. The presenter gives a broad estimate that laser material processing accounts for a substantial share—around 60–80%—of material-processing applications in EV manufacturing.

Speakers and sources featured

• Prof. Pasquale Franciosa: Main lecturer; presents his research, methods, and case studies.
• Workshop host/moderator: Introduces the session, manages questions, and thanks the lecturer. The auto-generated subtitles do not reliably identify the host by name.
• Omra: Participant who briefly comments near the end; the subtitles identify the speaker as “Omra,” but the name may be imperfectly transcribed.
• Sources and research discussed: Franciosa’s WMG research and collaborative case studies; published research papers and experimental results shown during the lecture; and public statistics about EV manufacturing and digital-technology adoption. No additional authors or specific publication titles are clearly identified in the subtitles.