Process equipment for solids processing (SP) obeys very specific rules of behavior. Many years ago, I was part of a team that wrote a program to make the selection of SP equipment easy. It was an early attempt at AI, an expert system based on rules. The first version of the program started by defining the starting materials and setting an objective. For example: our objective could be to produce a dry cake with a specific solvent residual level, starting with a slurry. Sounds easy.
That was when we discovered how little we knew about drying. The culprit was data. A few years later, a group of engineers and chemists developed a trouble-shooting guide for crystallizers, a much simpler topic. Our hope was that this group's knowledge could be harnessed in an expert system.
We started by classifying the types of crystallizers and defining their characteristics. However, we found that the chemical properties overshadowed the characteristics of the crystallizer. Often, the product was developed in a specific device (i.e., a mechanically mixed vessel) that hides some of the chemical’s properties. The steps taken by the chemist may have required more time, induced product changes, and involved much more expensive equipment than the product’s profit margin could sustain. Also, scale-up could result in a different physical or chemical form. We then needed to work with the chemist to identify the physical properties and extrinsic conditions that control the various steps in the process to convert the laboratory findings into a meaningful process design.
As before, our attempt at an expert system was quickly abandoned because of a lack of basic data.
Chemical Properties vs. Equipment Design
However, we discovered that trouble-shooting a batch or continuous crystallizer had similar problems caused mainly by the chemical we were making. Rather than solve a problem by deduction, working from general to specific, we worked from specific fundamentals such as nucleation and growth.
For example, a product made up of small crystals would point to excessive nucleation. That may suggest that it was caused by attrition, so our first step, after getting a solubility curve, was to look at the crystals. Getting nucleation rates can be very difficult to obtain, but observing the crystal shape and surface is simple and effective in resolving a supersaturation issue from attrition. These observations may identify other characteristics, such as agglomeration, that would help with the process design.
Our expert system broke down the problems into known issues with nucleation and growth and offered suggestions of specific data that would identify the solution to the problem.
This was homework based on observations rather than an experimental design, which can be expensive. We could then acquire basic information such as a solubility curve, polymorphs or enantiomers, meta-stable zones and the potential for generating supersaturation by other means (i.e., drowning out). On the equipment side, we needed the general arrangement for the crystallizer, how supersaturation is achieved, how product is removed, and the use of any external loops to enhance particle-size distributions.
Solving Problems Through Observation
In the example, the solute may be outside the meta-stable zone because of poor mixing. This could be due to the location of the feed or, in the case of a batch crystallizer, the rate of cooling or evaporation. Continuous crystallizers are not immune to this problem, especially if they have external circulation loops. The solute may have polymorphs that are more stable in a temperature range that favors a specific phase. Once these form nuclei, growth may remove solute from the solution and prevent the formation of the more stable polymorph.
Armed with this information, we took the Aristotle approach to our expert system, which was to work from these specific characteristics to general conclusions by induction. It allowed us to build an expert system that included not only physical properties of the intended product but also the cost of getting additional data or making changes to the system. The obvious first step was getting a solubility curve.
It is not unusual to find a process that has been in operation for several years without having a solubility curve. When we talked to equipment manufacturers, we were not surprised that they rely on past experience with a chemical in a specific device. This takes a lot of choices off the table, and the expense of determining more physical properties is often not warranted.
However, some of the properties needed can be cost-effective and simple to identify, such as clumping, flowability and agglomeration. I recall a process that could produce very large particles, but it took a day to make them. However, 50 to 100-micron particles could be generated in six to eight hours. After a few more discussions and experiments by our chemists, we looked at four alternative routes:
- Crystallize to 50-micron, filter and wash to remove impurities, and then extrude to 100 microns or larger. Drying produced very little dust.
- Feed the wet cake from above into a granulating fluid-bed dryer to get a 100-micron particle with a slightly wider particle size distribution.
- Dry the wet cake in a flash dryer and then spheronize the fine particles to about 100 microns in a tilted-pan dryer with water added back to agglomerate the particles.
- Crystallize to 100-micron, filter, wash and dry as originally planned, but use the fines from downstream processing as seeds for the crystallizer. This reduced the crystallization time to less than 2 hours.
So, our expert system didn’t really solve any specific problem, but it did open our minds to alternative solutions to the system. I hope AI will do the same with the improvement in processor speed and larger databases that have consistent fundamentals.
