**Machine learning, a type of artificial intelligence, is becoming increasingly prevalent in everyday life. Email spam filters, autonomous cars, and speech recognition all rely on machine learning algorithms to function accurately and efficiently. Such algorithms allow computers to find trends in data without explicit programming or problem awareness, allowing them to make predictions when exposed to new inputs. **

** **

**This paper explores the potential of this technology to assist in the field of structural glass design. Supervised regression multilayer neural networks and classification algorithms are trained on a database of computational structural glass solutions generated parametrically in Grasshopper and Strand7. Once trained, the algorithm’s accuracy is assessed and used to predict glass build-ups for a rectangular plate with uniform pressure. **

** **

**Thickness classification accuracies of greater than 80% are achieved in all cases. When used in combination with experienced structural engineers, such intelligent predictors have the potential to offer benefits in early stage design, allowing rapid and accurate assessment of glass and consideration of the wide variety of design drivers involved in structural glass design. **

**Introduction**

** **

As with other technical fields, fundamental engineering problems can be formulated and automated based on their mathematical principles. The principles of computer aided design (CAD) have long being outdated and we have progressed into the domain of computer “generated” design. As designers and engineers we are increasingly more reliant on computational power not only to solve technical problems but to generate the designs in the first place. Architects and designers resort to computational design as a means of exploring a multitude of solutions which are sub-optimal with increasingly wider range of parameters, including solar, thermal, structural, views and aesthetics.

Currently, despite the computer advances, expertise in design and engineering has been provided by individual human input. As a result most problems are solved through the prism of personal experience and judgment. In this human experience, learning is critical. While this works well on an individual basis, the power of the collective experience cannot be mobilised due to the knowledge having to be passed on. With this emphasis on the quality of learning, we flag the importance of learning as a principle in digital design. Furthermore, the principle of computer learning can be applied to any task for which basic parameter data can be gathered and stored in a logical way.

Machine learning is the definition of a series of algorithms originating in computer science and mathematics, specifically in the subset of artificial intelligence. For the purposes of applied engineering it can be described as self-teaching algorithms that are trained on a set of data gathered in the course of a process. The algorithms make future predictions by extrapolating from past-experience. For welldefined, trivial engineering problems there may be little benefit in their application, but the algorithms have unparalleled potential to solve very complex problems.

As our problems become more reliant on numerical analysis and integration of disciplines, the separation of individual aspects of a problem will become more difficult. As we evolve further to solve those complex problems entirely in the digital domain we will need to classify, record and discretise as much of the input and output information as possible. Machine learning uses this information, stored in large organised databases to create predictive mathematical models based on the analysis of this data.

In Figure 1 the data for a two dimensional problem can be analysed and a function can be fitted to represent the data based on a set of criteria.

**Figure 1: Curve fitting to a set of data points**

**Applied Machine Learning in Other Industries**

Two transferable examples are highlighted below illustrating how applying machine learning could offer value to companies and clients throughout the design chain of the glass industry.

I. Google DeepMind: Minimising data centre cooling loads

By harvesting data collected from thousands of sensors within their data centres, Google DeepMind trained neural networks to predict energy use, future temperatures and future pressures in their data centres. When run on live data centres, this machine learning system was able to consistently demonstrate a 40% reduction in the energy used for cooling [1].

II. Otto: Predictive inventory management Otto, a German based e-commerce company, have trained a deep learning algorithm to predict what customers will purchase a before they order. By analysing 3 billion past transactions and over 200 variables, the algorithm is able to predict with 90% accuracy what will be sold in the next 30 days. It purchases around 200,000 items automatically each month with no human intervention. As a result, Otto’s surplus stock has reduced by one fifth, customers get their products sooner, and product returns have reduced by over 2 million items per year [2].

**Methodology. Problem Outline **

While machine learning is successfully used to solve complex optimisation problems, its use in the field of structural and façade engineering is not yet explored. The work presented here aims to prove the suitability of this technique in the field of structural glass. A common design task in structural glass is to find the minimum thickness of a four-side supported rectangular plate for given stress and deflection limits. This problem is well understood, with analytical solutions documented in Rourk [3]. As such, it provides a well bounded, restricted domain, comparative example to illustrate and validate the suitability of machine learning methods for this particular design situation. The complexity of the problem can be generalised to include more parameters and constraints in the future. The problem outline is: “Predict** thickness** for a given** width, height, pressure** and **glass type.**”

* *

(...)

**James Griffith, Vladimir Marinov, Giulio Antonutto**

Arup

**References**

[1] Evans, R. and Gao, J. (July 2016) DeepMind AI Reduces Google Data Centre Cooling Bill by 40% [online] Available at: https://deepmind.com/blog/deepmindai-reduces-google-data-centre-coolingbill-40 [Accessed 3 May 2017]

[2] The Economist (April 2017) How Germany’s Otto uses artificial intelligence [online & print] Available at: http://www.economist.com/news/business/21720675-firmusing-algorithm-designedcern-laboratory-how-germanys-otto-uses [Accessed 3 May 2017]

[3] Young, W. and Budynas, R. (2011) Roark’s Formulas for Stress and Strain 8th Ed. McGraw Hill Companies

[4] Mirtschin, J. (2017) GeometryGym [online] Available at: https://geometrygym.wordpress.com/ [Accessed 3 May 2017]

[5] Matlab (2017) Neural Network Toolbox [online] Available at: https://uk.mathworks.com/products/neural-network.html [Accessed 3 May 2017]

The article is based on the lecture presented at

the GLASS PERFORMANCE DAYS 2017 Conference,

which took place on June 28-30, 2017

in Tampere, Finland

*Całość artykułu w wydaniu ***drukowanym ***i** ***elektronicznym** *Inne artykuły o podobnej tematyce patrz* **Serwisy Tematyczne** *Więcej informacji:* **Świat Szkła 02/2018**

** **