Update 30th Sept 2025:

• We have ca. 350 registrants, so it seems like we will have a similar or greater turnout than last year. 

• Make sure you get your discussion round requestion submitted in time here.

• Here is inspiration for questions from last year: https://www.active-learning.uk/2024

Expected Outcomes and Benefits

• Meet with academic experts and industry innovators in a focused symposium to hear the latest case studies in the use of Active Learning/Bayesian Optimisation in a variety of industry applications including pharma, semiconductors, chemistry and materials.

• Stay informed about the latest developments. This programme takes a deep dive into best practice approaches and latest technological advancements; bringing together thought leaders for focused discussions and networking. Discuss and facilitate knowledge exchange about the latest use-cases.

• Understand current challenges and limitations. What challenges will I face in the future? What do I have to prepare for? Where are my strengths and weaknesses, compared to other teams and their challenges?

Active Learning 2025 Program

(2pm London, GMT)

Warmup: 13:00 Crash-Course: Bayesian Optimization for Physical Experiments (For Beginners) by Morten Nielsen

• 45 minute Tutorial: Motivation, Usage, Example

• 15 minute QnA

------------------------------------------------- 

14:00 15-minute Keynote [Main Start]

1. Setting the scene: The 2025 Active Learning Challenges 

2. Paying it forward: Meet the 2025 Active Learning Grant holders (Gideon, Dennis from Ghana)

3. Making the most of it: Your contribution matters

14:10 "Meet the Teams": Introductions

1. Roche: Joshua W. Sin

2. Takeda: Andrew Kukor

3.  Vertical Cloud Lab: Sterling Baird

4.  Benchling: Janet Matsen

5.  Interested to present your Active Learning/BayesOpt team? Get int touch by end 14th November.

14:30: "Stories from the Frontlines": (Un)successful Case Studies  in Active Learning

6. Evotec: Iterative Screening - Unlocking High Throughput Screening Opportunities for Challenging Assays (Benedikt Bauer)

   • Abstract: "Iterative Screening combines machine learning-guided compound selection with physical screening in successive small batches. Unlike classical High-Throughput Screening (HTS), this approach leverages predictive modeling to prioritize compounds based on prior screening outcomes, resulting in significantly higher hit rates while reducing the number of compounds tested. The iterative nature of the process allows for dynamic refinement of compound selection, enabling more targeted and efficient exploration of chemical space. This methodology is particularly advantageous in assays that are complex, resource-intensive, or constrained by limited material availability. In such settings, Iterative Screening has the potential to achieve up to a fivefold increase in hit rates compared to traditional HTS. Overall, it offers a more adaptive, cost-effective, and data-driven strategy for Hit ID."

7. Monomer Bio:  24-Hour Mystery Cell Culture Challenge (Jimmy Sastra & Yang Choo)

   • Abstract: Lessons from a 24 hour hackathon on building closed-loop lab automation systems.

8. Merck KGaA: How BayBE revolutionizes iterative planning at Merck KGaA (Alexander Hopp)

   • Abstract: BayBE (https://github.com/emdgroup/baybe) is an innovative, open-source framework for Bayesian Optimization developed by Merck KGaA, Darmstadt, Germany, that revolutionizes the way researchers approach experimental design and optimization. It enables scientists to streamline their workflows, reduce resource consumption, and accelerate product development cycles. In this talk, we will provide a brief introduction to BayBE and discuss some of the key use cases we have encountered at our company. By highlighting these examples, we will showcase how BayBE enhances the efficiency and effectiveness of experimental planning, ultimately driving innovation and collaboration within our research teams.

------------------------------------------------- 

15:00 Discussion Round: pre-defined questions first, then open questions

1. Technical

   1. When are custom surrogate models useful? When are they overkill?

   2. What are the trade-offs when using batched acquisition functions? Are there batches 'too big'?

      1. Are there specific tricks to batched AL with batch sizes of 96 or more?

   3. How close are we to having good theory (or heuristics) for when/where AL methods will be more effective than random search? (random search can be surprisingly effective)

   4. How can/should active learning be done in the era of foundation models? Eg should foundation models _replace_ active learning? Are there some problems better for 1 type than another?

   5. What are the right research questions to improve multi-modal / multi-fidelity active learning?

2. Ecosystem

   1. What are the best Bayesian Optimisation packages at the moment?

   2. How does the setup of active learning problems in academic papers differ from what real world users want?

3. Adoption

   1. How do we measure the efficacy of Active Learning e.g. versus status quo?

   2. Active learning seems so rarely implemented in industry. What are the biggest barriers to adoption?

---------------------------------------------- 

16:00 Cutting Edge: Industry Talks 15 mins each, x2

1. Challenges and Opportunities in Cutting-Edge BO Methods (Jordan Penn)

   1. Abstract: TBD

2. Robochem-Flex: Democratizing Chemical Optimization (Elia Savino)

   1. Abstract: Robochem-Flex is a modular, 3D-printed platform that brings self-driving chemistry within reach of any lab. By combining low-cost hardware with advanced Bayesian Optimization, it enables fully automated reaction discovery and optimization without the need for specialized robotics. Its flexibility allows rapid reconfiguration for diverse chemistries, while its affordability democratizes access to self-driving lab technology. This talk will showcase how Robochem-Flex bridges the gap between cutting-edge automation and everyday chemical research.

16:30 Cutting Edge: Academic Talks 15 mins each, x2

1. Posterior Sampling for a Broad Class of Experimental Design Tasks (Raul Astudillo)

   1. Abstract: Many experimental design tasks go beyond optimization. Scientists may need to map feasible regions, identify level sets, or select designs that satisfy multiple criteria. These complex goals call for adaptive algorithms that are both general and efficient—yet existing methods are often difficult to adapt to such settings. In this talk, I will present an algorithm that generalizes posterior sampling (Thompson sampling) to a broad class of experimental design goals. The method is simple, scalable, and naturally parallelizable, while performing competitively with far more computationally demanding information-theoretic approaches.I will show how this perspective unifies diverse adaptive design tasks within a single framework, highlight its empirical effectiveness across applications, and discuss theoretical guarantees that position posterior sampling as a powerful paradigm for experimental design.

2. Constrained Composite BO for By-design Particle Synthesis (Fanjin Wang)

   1. Abstract: People often use Bayesian optimisation to search for 'best' materials, drug formulations, processing conditions... but what do we actually mean by 'best'? In most cases, this 'best' means maximising, or minimising a property. But what if we want to hit a predefined target? In biomaterials research, a hydrogel may need to match physiological mechanical properties - not just be as firm as possible. In drug development, a formulation is developed with precise release profile to achieve desirable pharmacokinetic and pharmacodynamic effects. This work deals with the target-value optimisation problem - further complicated by a priori unknown experiment constraints in nanomedicine. The approach not only outperform all tested BO frameworks, but also beats 14 wet-lab experiment experts in a nanoparticle synthesis optimisation task."

------------------------------------------------- 

17:00 Startup Corner: 

Reactwise: Data-Driven Transfer Learning for Fast Reaction Optimization (Daniel Wigh)

Abstract: Optimization of chemical reactions remains a major bottleneck in process development, where each new system often requires extensive experimentation from scratch. Despite the growing availability of reaction data, effectively reusing prior knowledge across related systems remains a challenge - naive transfer can slow learning or even degrade performance. 

At ReactWise, we generate large-scale proprietary experimental datasets in our own lab across key reaction classes - amide couplings, Buchwald–Hartwig couplings, and Suzuki–Miyaura couplings - to investigate how prior chemical knowledge can accelerate optimization. I’ll present our work on transfer learning in Bayesian optimization, comparing two approaches: a multi-task Gaussian Process and a Ranked Gaussian Process Ensemble (RGPE). RGPE models introduce a data-dependent weighting decay that dynamically adjusts how much prior data to trust, enabling rapid adaptation to new systems while avoiding negative transfer, while multi-task GPs tend to be better suited when tasks are highly similar. 

Using thousands of high-throughput experiments, we demonstrate that informed transfer learning can cut optimization times by over 50%. This work points toward a future of generalizable and data-efficient optimization with our proprietary transfer learning algorithm MemoryBO, where high-throughput experimentation and adaptive algorithms jointly enable “memory-driven” process development."

Saddlepoint Labs: a connected system of sensors and cameras to provide intelligence throughout the laboratory (Peter Frazier)

Abstract: Laboratory automation provides throughput and repeatability but requires significant investment and is difficult to adapt to new tasks. Humans are flexible but prone to cognitive lapses and technique variation that limits data quality.

Rather than replacing humans, Saddlepoint Labs aims to make them better. We build vision-based co-bots that provide experimentalists with a second set of eyes --- using computer vision to detect errors, provide guidance, and record procedural data for post-experiment analysis.

Our current product supports manual pipetting, tracking which wells are used and precisely when each action occurs. This foundational data —-- exact timing and confirmed well locations—addresses a critical gap: the procedural metadata that active learning and quality assurance systems need but experimentalists rarely record. By automatically capturing execution details, we reveal errors, technique variations, and timing patterns that drive experimental variability.

Future systems will monitor other laboratory procedures and prevent activities that endanger safety or risk contamination. This human-in-the-loop approach preserves researcher expertise and flexibility while improving experimental reproducibility and data quality. For labs running active learning workflows, better execution monitoring means each carefully-designed experiment is executed as intended, with trustworthy data to train the next iteration of models.

Causal Chambers:  Empowering Machine Learning research with physical testbeds (Juan Gamella)

Abstract: We provide self-contained physical laboratories where researchers can safely test their algorithms in a real, non-simulated environment. We will give a short overview of our different labs (a.k.a. chambers), how they have been used in different fields of ML & stats, and how you can run your own experiments in real-time using our remote API.

------------------------------------------------- 

17:20 Conclusion and 2026 Symposium Announcement