ÐÓ°ÉÔ­´´

Medieval swordsmiths knew little about the physics or chemistry of the
metals that they repeatedly hammered, heated and quenched. What they did
have was empirical knowledge handed down from master to apprentice and
perhaps slowly refined over time. Progress in the intervening centuries has
not been dramatic. The computers that now run many materials processing
operations still rely mainly on empirical information; they are programmed
to maintain a set of conditions, such as temperature and pressure. While
this approach has had remarkable success, it is limited by its dependence on
experience.

Overcoming such limitations is the goal of the new technology of
intelligent, or knowledge-based, processing of materials. At the heart of
this technique are mathematical models that relate a material’s
properties, such as strength and flexibility, to its internal structure.
The models also seek to explain how this structure depends on the conditions
under which the material is produced, and so help to control the
manufacturing process more accurately to boost productivity and, ultimately,
make superior products.

Devising models that apply to a broad class of materials will reduce the
need for time-consuming and costly laboratory experiments every time a new
process emerges. By fostering a better understanding of a material, how it
can be made and how it behaves, intelligent processing should also make it
possible to control much more accurately.

The activities to which intelligent processing can be applied fall into two
broad classes: the manufacture of commodities, such as steel, aluminium,
glass, cement and polymers, and the fabrication of speciality products, such
as silicon or gallium arsenide for the computer industry, composite
components for aircraft, and optical fibres for telecommunications.
Commodities are produced in quantities measured in hundreds of thousands or
even millions of tonnes each year. The plant needed to manufacture them is
costly, its operation tends to be inflexible, and long-term changes are
difficult to implement. For example, the basic technologies used to produce
steel, glass and aluminium have remained substantially unchanged over many
decades. Productivity is the critical issue – arranging the production line
to minimise the costs of energy, labour and raw materials. There is little
incentive to improve quality as long as a basic requirement is met.

With speciality materials, production is on a much smaller scale, though the
monetary value is comparable: a single wafer of gallium arsenide 20
centimetres across is worth about as much as a tonne of finished steel.
Because quality and performance are all-important, engineers continually
develop new processing techniques for these materials. The capital
investment required is much smaller than in industries producing the
commodity materials, so changes in equipment are likely to be relatively
frequent.

Intelligent processing can benefit both segments of the materials industry.
For the commodity industries, intelligent processing should improve
efficiency and so cut costs. For high-technology materials, the
knowledge-based approach might yield new production methods and, possibly,
novel materials.

All process-control arrangements, from a household thermostat to a guidance
system for a smart bomb, work in basically the same way. Sensors monitor
conditions such as temperature, pressure and chemical composition, and the
process is adjusted to correct any discrepancy between these measured values
and some predetermined desired settings. In intelligent processing of
materials, the information obtained by sensors is not used directly.
Instead, it is processed or manipulated by mathematical models that perform
calculations based on equations that describe the process and the material’s
structure. It is the results of these manipulations, rather than the raw
output from the sensors, that controls the process.

In any materials processing operations, a fluid (either a gas or a liquid)
is transformed into a solid of the desired shape and structure. The nature
and speed of chemical changes that occur and the dynamics of fluid flow and
heat transfer all affect the way this transformation takes place. To control
the outcome it is necessary to know about these changes, and to understand
the relationship between a material’s internal structure and its properties.
This kind of information is often either not available or only partially
used.

Consider, for example, what happens when two pieces of metal are welded.
This is a complex process that involves the transfer of heat from a welding
arc, melting, circulation within the molten material, and the heating of a
much larger region around it. The welder is probably not aware of these
phenomena at this fundamental level, but will have accumulated a great deal
of practical knowledge of how to make a good weld. By contrast, the
technical specification for the welding operation will typically have been
set by engineers, with little input from the operators themselves.

Robots have taken over from human welders in the car industry, among others.
But these machines merely reproduce the actions of their skilled human
counterparts, and do so in a way that is suited only to a limited set of
circumstances. For every change in the size of the workpiece, and for every
new material, the robot has to be reprogrammed, and this is costly. An
intelligent welding machine, by contrast, would continuously monitor the
temperature of various key points in the area being welded. A mathematical
model would compute from these readings critical information such as the
depth of the area and the thermal stress on the materials being joined. The
results of these computations could then guide the robot’s work – perhaps
making it go faster or more slowly, or deliver heat in patterned bursts
rather than continuously.

The harsh environment of many processing operations has limited the use of
intelligent control in the materials industry. Steel is made at around 1600
°C, while silicon melts at around 1420 °C, conditions that
make it difficult to measure temperatures and monitor composition
continuously.

Recognising the potential of intelligent processing, the US Department of
Energy and the American Iron and Steel Institute have combined forces in a
$50 million programme to develop sensing and control strategies for
the steel industry. In one project, the Los Alamos National Laboratory and
the Pennsylvania-based corporation Bethlehem Steel are collaborating to
develop a fibre-optic system to monitor temperatures in a basic oxygen
furnace. The fibres transmit visible and infrared light from the furnace to
sensors a safe distance away, and from the distribution of energy in the
spectrum the system computes the temperature of the furnace. This allows
continuous monitoring of the furnace, and should lead to much closer process
control.

Deductions at a distance

It is not just for measuring temperature that remote sensing technologies
are being developed. Magnetic sensors can detect the level of molten metal
in a closed container, for example, and lasers can measure an object’s shape
or dimensions. So-called eddy current sensors induce a current in a
specimen, and from the resulting magnetic field detect surface defects and
changes in the porosity of the material.

Even with such methods, some of the most important characteristics of a
material are difficult to sense directly during processing. Take a
material’s microstructure, for example. Is it crystalline or amorphous, and
if it is crystalline what is the atomic arrangement and how large are the
crystal grains? But when the measurements are processed by mathematical
models, such information can be deduced.

There are two basic types of mathematical models: empirical models and
mechanistic models, and both are important in intelligent materials
processing.

Empirical models are compiled from experimental observations. Take, for
example, the transformation of steel slabs 200 millimetres thick into sheets
just 1 millimetre thick. A rolling mill requires several stages to produce a
sheet of the specified thickness and with the desired microstructure and
properties. A model based on accumulated experience determines the number of
times the steel passes through the mill, the reduction per pass, and the
temperature at which this deformation has to take place to yield acceptable
quality.

Model metals

But there is no guarantee that experience has ever achieved the best
conditions. In rolling, a four-pass process might have worked better than a
two-step process – but clearly there are many more possibilities. Perhaps 7
or even 17 steps might prove even better. Or maybe temperatures should
increase by 10 °C during each rolling operation. It would be much too
costly to conduct such wide-ranging experiments for every material.

Mechanistic models can help supply this information. These models are based
on fundamental chemical and physical principles, such as Newton’s laws of
motion and the theory of chemical equilibrium, that govern the behaviour of
fluids and solids. The complicated differential equations that embody these
laws can be solved by computers to yield a representation of many physical
phenomena that are based on fundamental principles. Armed with this
knowledge it is possible to calculate accurately the distribution of
temperatures, velocities and magnetic fields in molten materials, their
rates of melting and solidification, and the mechanical stresses that they
experience under a variety of conditions.

By sharpening our fundamental understanding of a process, a mechanistic
model enhances the ability to control it. Fundamental analysis of steel
making, for example, suggested that intense agitation would accelerate the
transformation of molten iron into steel. Modern steel plants apply this
insight to accomplish in 30 minutes what used to take 12 hours. In steel
casting, modelling may allow us to improve significantly the quality of the
solidified metal. It might, for example, suggest that strong stirring, or
agitating the metal while it freezes, should have the desired effect. In
other processes, keeping the melt essentially motionless in a strong
electromagnetic field may yield a material with the desired qualities.
Modelling provides quantitative understanding of these phenomena, which in
turn indicates the direction to head in when experimenting with new or
improved processes.

In the electronics industry, for example, reliable soldered connections are
crucial. Attaching a single microprocessor to an electronic circuit board
can require several hundred connections. Yet even the straightforward
question of what shape a molten solder bead tends to take when making the
connections has gone unanswered for many years. At the Massachusetts
Institute of Technology, materials scientists recently developed a computer
model that calculates the complex shapes of these beads. This information
should permit designers to decide more precisely how much solder is needed
to provide a reliable connection and how closely spaced the connections can
be without risking short circuits between them.

Although the output of mathematical models is numerical, the insights gained
from them are often qualitative. A good example is the formation of
metallic or ceramic coatings by spraying molten droplets onto a solid or
molten surface. A quantitative understanding of how these droplets flatten
or spread on impact, and how the impact velocity affects the spreading and
adhesion process has led to the development of an entirely new technology.
This is the high-velocity oxygen fuel process, where droplets are driven
onto the substrate at supersonic speed, yielding better coatings than those
produced by conventional methods.

Intelligent processing has so far been applied mostly to the production of
commodity materials. In Japan, several steel companies, including Nippon
Steel, Kawasaki Steel, Kobe and Sumitomo, are now operating large
computer-controlled blast furnaces. Iron ore, coke and flux materials are
fed into the top of furnaces up to 10 metres in diameter and 30 metres
high, while hot air enters at the bottom. The solid material gradually
descends, reacts and melts, yielding hot pig iron and molten slag that is
continuously tapped at the bottom. A typical blast furnace may produce up
to 10 000 tonnes of molten iron per day, valued at $1.5 million,
equivalent to more than $500 million a year.

The main issues in blast furnace operation are the efficient use of coke,
which is the most costly of the raw materials, and the quality of the iron
produced. To achieve the best possible operating conditions, Japanese
operators fitted the blast furnace with numerous sensors that measure the
temperature and composition of the gas given off and correlate readings
from them with the quality of the product and the amount of raw material
used. The steel companies also use a great deal of fundamental mathematical
modelling to represent the chemistry and physics of the process, including
reaction rates, fluid flow phenomena and heat transfer, and by these means
they have developed a set of rules that enables a computer to control the
blast furnace efficiently. The payoff has been substantial: intelligently
controlled blast furnaces may consume between 5 and 20 per cent less coke
than conventional systems, with corresponding cost savings.

Although still well behind the Japanese, the US steel industry is beginning
to turn to intelligent processing. Some American rolling mills and
continuous casting operations use intelligent processing to improve the
microstructure of the metal and the quality of its surface. This reduces the
number of substandard pieces turned out, and saves energy and labour – all
factors implicated in US producers’ sharp loss of market share during the
1970s and 1980s. In recent years they have become competitive, and regained
some of the lost ground.

In the use of advanced technology in the US, the so-called minimills are
taking the lead. Chaparall Steel, a successful minimill operator in
Midlothian, Texas, relies on intelligent processing to optimise its electric
arc furnace systems. Chaparall uses an electric arc of up to 100 megawatts
to melt and refine scrap steel, which is then formed into components. Arc
furnaces offer important environmental advantages, but they are still far
from being optimised. It is difficult to maintain a stable arc that doesn’t
‘flare’, and the electrodes wear out quickly. Chaparall Steel adopted a
computer control technique which originated in Mexico. It promotes stability
by carefully balancing the phases of the electrical current and the
positions of the electrodes, and has led to appreciable energy savings. A
similar system is being developed in the US by Milltech-OH of Davenport,
Iowa.

Human hurdles

While the physical chemistry of steel making and the flow phenomena in the
iron blast furnace have been studied for decades and are thoroughly
understood, many high-technology operations still defy precise scientific
understanding. Semiconductor fabrication is one example. Single crystals of
silicon or gallium arsenide are formed by slowly pulling them out of a melt
contained in a cylindrical, rotating crucible. The crystals, between 50 and
250 millimetres in diameter and more than a metre long, are then sliced into
thin wafers from which the chips are fabricated. Producing high-quality
crystals requires precise control of the circulation and temperature within
the crystal and melt, and of the rate at which the crystal is pulled.

While a great deal of information has been collected on this process over
the years, the systems are still designed and controlled largely on an
empirical basis, and optimal process conditions are not always easily
defined. The best results come when part of the melt is well stirred, to
produce uniform composition, while the fluid near to where the crystal is
being drawn out is kept quiescent, to avoid disrupting the fine structure of
the solid. This can be achieved by applying a rapidly fluctuating magnetic
field in the bulk of the melt and a steady field near the crystal surface, a
method I recently patented. A better understanding of the basic physics of
the process could lead to more efficient production and higher-quality
material.

Some of the most noteworthy work directed towards intelligent processing of
high-technology materials is being done at the National Institute of
Standards and Technology in Gaithersburg, Maryland. In one project, the NIST
is working on the production of fine metallic powders used in the
manufacture of components for cars and aircraft, such as landing gear,
turbine discs and engine superchargers. A stream of molten metal is broken
into droplets by a high-velocity jet of gas. The droplets then solidify to
yield the powder. The interaction between the gas and the liquid must be
precisely controlled if the powder particles are to have the uniform size
needed to produce high-quality finished components.

In conventional processing, the powder produced is periodically analysed,
and the process is adjusted according to the results. But the time lag
between analysis and corrective action means that large quantities of
substandard powder are often produced before the system is readjusted. In
the NIST scheme, laser diffraction systems continuously analyse particle
size to provide data that are fed into a model that relates the process
conditions to the powder quality. This allows the pressure and velocity
distribution in the gas jets to be adjusted ‘on the fly’, yielding a
uniformly sized powder.

Perhaps the greatest potential for intelligent processing lies in developing
entirely new process technologies, something that is at present left largely
to experts’ intuition. Models can serve a kind of screening function –
determining, for example, whether a proposed new process would conform to
the basic laws of chemistry and physics. Some of these laws are obvious,
such as the need to conserve matter and energy, but other constraints can
be more subtle. It is important, for instance, to understand the rates of
competing chemical reactions, to assess whether specimens can be heated or
cooled rapidly enough to produce the desired outcome, and to determine
whether a material’s surface will remain stable. Mathematical modelling may
also reduce the need for experimental work on new processes by helping to
establish the ideal set of conditions, such as temperature, pressure and
patterns of fluid movement.

Although the concept underlying intelligent materials processing may seem
like common sense, there are serious human barriers to its wider use. The
materials scientists engaged in product development are not usually
comfortable with the engineering tools of sensing, control and database
development. Conversely, the engineers who are expert at process control
tend not to be knowledgeable about materials science.

The two approaches to modelling are separated by a similar gulf. Mechanistic
models emanate from laboratories that are used to dealing with the basic
problems of engineering, while empirical models are built from accumulated
industrial experience in factories, steel mills or semiconductor fabrication
plants. While scientists are continually sharpening their quantitative
understanding of the fundamental properties of the materials, this
knowledge does not flow quickly from the laboratory to the shop floor.

The concept of intelligent processing originated in the US, but it is small
wonder that the Japanese, who excel at forming interdisciplinary teams, are
pushing the technology the hardest. The rest of the world should follow
their example as master swordsmiths of the modern era.

Julian Szekely is professor of materials engineering at the Massachusetts
Institute of Technology and a member of the US National Academy of
Engineering. This article first appeared in the May/June 1993 issue of
Technology Review, which is published eight times a year by MIT. Copyright
by Julian Szekely/Technology Review. Distributed by New York Times
Syndication Sales.