A neural network has been trained to predict the outcome of a chemical reaction controlled by molar ratios, temperature, pressure, amount of enzyme and stirring speed.Kirk, Barfoed, and Bjorkling at NOVO Nordisk A/S in Denmark used the BrainMaker program to train their neural network to predict the amount of desired product and by-product which would be formed after 22 hours of reaction time.
An excellent correlation between predicted yields and experimental results was found. The neural network saves time and money by predicting the results of chemical reactions so that the most promising conditions can then be verified in the lab, rather than performing a large number of experiments to gain the same information.
Initially 16 experiments were performed to identify the most important parameters controlling the process. The molar ratio between fatty acid and glucoside, reaction temperature, pressure, amount of enzyme, and stirring speed were varied. The synthesis yielded ethyl 6-O-dodecanoyl D-glucopyranoside. This experimental data was used to train the neural network to output the amount of the 6-O monoester and a diester by-product, represented as a percentage of yield.
The neural network had three layers: 5 input layer neurons, 4 hidden layer neurons, and 2 output layer neurons. It was trained using the back propagation algorithm with the sigmoid threshold neuron function.Twelve facts were used to train the network to an accuracy of 96% for the outputs. In only a few minutes, all facts were learned. The trained network was then asked to make four predictions on data it hadn't seen before. The network predictions were compared to experimental observations. Very good correlations were found.The average deviation between the network and the experiments was 4% (percentage of yield), ranging between 2% and 7% difference. These deviations are within the normal experimental error of synthesis.
After being tested, the network was put to work evaluating thousands of possible conditions in order to find the most optimum.Using a simple algorithm, a test file was generated containing all of the possible values, totalling 9900 cases.The computer- generated test file contained values for each parameter which were both within and without of the training value's range. The entire file ran through the network in 7 minutes and the predictions were saved in a file. Using a search function, predictions for specified yields were selected. Only three cases were found to predict more than 88% monoester with a less than 4% formation of the diester. One of these cases was tested in the lab and the results were close to experimental observation. The network had predicted 88.1% monoester and the experiment yielded 86.2%. The network predicted 4.0% diester, the experiment yielded 4.8%.
Finally, the 9900 predictions were again searched, but this time with additional restrictions more suitable for large-scale chemical processing. Again, the experimental results were very close to the yields predicted by the network.
 "Application of a Neural Network in the Optimization of an Enzymatic Synthesis," Ole Kirk, Martin Barfoed, and Frederik Bjorkling, NOVO Nordisk A/S, Novo Alle, DK-2880 Bagsvaerd, Denmark.